Projects
![]()
No projects found
Investors
New
DAO Treasuries
DATs
Lend / Borrow APRs
Chains
Categories
Spaces
How DeFAI agents can outperfom traditional yield strategies
Recorded: Aug. 6, 2025 Duration: 1:06:51Space Recording
Short Summary
In a recent discussion, industry leaders from Gauntlet, Giza, Almanac, and Brahma explored the transformative potential of AI agents in DeFi, highlighting their role in optimizing yield strategies and driving growth in the sector. The conversation underscored the importance of partnerships and the need for human oversight in AI-driven financial decision-making, while also addressing emerging trends in personalized finance.
Full Transcription
Thank you. Thank you. Thank you. Thank you. . Hi, hi, hi, everyone.
Welcome to the Spaces.
We're just waiting for one more speaker to join and then we'll keep things off probably
on time actually, which is fantastic. . And we are starting right on time, which is phenomenal.
Welcome, everyone, to this space is hosted by Gauntlet on how DeFi AI agents cannot perform traditional yield strategies.
My name is Simon.
I'm the product marketing lead here at Gauntlet working on yield strategy products. DeFi AI agents cannot perform traditional yield strategies. My name is Simon.
I'm the product marketing lead here at Gauntlet working on yield strategy products.
And I am super excited because I'm personally super bullish on AI agents.
I'm really excited about what's being built in the space. I am joined here by Wrench from Giza, Arma agents, Neo from Almanac, and Alessandro from Brahma.
I guess we can do like a quick round of introduction.
Renj, why don't you kick us off with a little intro?
Sure. Thank you so much for hosting us, Simon.
Pleasure to be here. Hello to everyone.
My name is Renj, I'm the CEO of Giza.
At Giza, we have an infrastructure for
automated finance and scale. And this infrastructure allows us to build financial autonomous
agents on top of it. We have built the first one in Web3 and we are looking to expand to many,
many other adventures. I'll pass it on to Neo from Almanac. Hey guys, can you hear me?
Neo from Almanac.
Hey guys, can you hear me?
Yes, we can.
Awesome. Pleasure to be here. Thank you very much.
So I'm Neo.
I'm the CEO and co-founder of Almanac.
And in Almanac, we like to call ourselves a vibe coding company.
We use AI agents to build fully deterministic, verifiable and auditable strategies.
So our agents build a code that doesn't evolve.
So basically we have copied a traditional workflow of traditional hedge funds and trading desks
and basically we replaced all the people that work there with agents.
agents. So you can come to the platform called the strategy and deploy the strategy. This
So you can come to the platform, code the strategy and deploy the strategy.
strategy can be backtested, simulated, audited. Everything, of course, is fully private, so
you don't share the strategy code. But we made this choice because we believe that if
you want to manage billions of dollars, you can rely on AI and you can rely on black box and NLMs. So that's about us and very excited to be here.
Thanks a lot, Neo. Alessandro, do you want to give a quick intro about yourself?
Hi, thanks for having us. I'm Ale, one of the co-founders at Brahma, and we have been focusing on
automation and orchestration for a long while now, and agents are part of it.
So not only we're able to programmatically issue accounts,
we're able to add security policies on top of them,
and then enable automated execution thanks to an order flow
and execution system that we provide to the agents.
And lately, we just launched cards as well, so credit cards that can be attached to the agents. And yeah, lately we just launched cards as well.
So credit cards that can be attached to the system.
And so down the line, we can also give off-chain payments,
connectors to both agents and bots to be able to perform actions,
such as payments in real life as well.
Awesome. Thanks a lot, everyone, for the intro.
I'm super happy to have you all three here with me. So
in terms of the agenda, what we're going to do is I've got a few questions to ask the three of you.
And then if we get a little bit of time at the end, we'll see. We can maybe get a couple of
questions from the listeners. But I guess we can just dive right into it. And the first question
is, I think everyone's like, everyone's asking that. Everyone who sees
DeFi x AI things is wondering this, which is like, why do we need AI agents in DeFi in the first
place? Ali, why don't you kick us off with this one? Yeah, I think there's a lot of potential
answers, but most importantly, wherever things are
in static and so they cannot change in an ongoing matter and more and more factors emerge,
agents do help in assessing opportunities with replacing some of the traditional people
or scripts that would be focusing on that as a peer mentioned.
So that's important.
And most importantly, both reasoning and acting
with a certain amount of latency,
especially when you take DeFi and how liquidity can change
really quickly or security risks can emerge,
it does make sense to go into essentially agentic reasoning
and let them all onto certain tools,
gather them to data and reason in a short amount of time.
Yeah, so these are very important.
And the important thing to attach to them
is some very strict guardrails to make sure
that they don't do more than they should.
Thanks, Ali.
Neo?
Thanks a lot.
Yeah, so I actually think we don't need agents in DeFi.
I think DeFi layer is great as it is. And the main advantage of DeFi layer
is that it's a financial system on blockchain.
What I think, and it's great, it's permissionless, it's safe, it's secured.
You keep money there and you feel that this money is saved there.
And I think whoever makes money knows that the most important thing
while making money is losing the money.
And if we add agents and AI to it,
the risk of losing money increases. So what we think at Almanac is that agents are great
to be used to basically discover DeFi.
So basically to reason about where is the best alpha,
where is the best opportunity and to build strategies.
So basically to execute those strategies as soon as possible, or maybe build those
strategies as soon as possible.
But making decisions about it should still stay for a person, for a human.
The person still should be responsible for that.
At Amalak, we are very pragmatic.
So we ask ourselves what AI is the
best app. And we know that AI agents are the best, or in general, AI is the best with reasoning. It's
one billion times better than a human being. Therefore, it can process data a billion times
faster. And it's better with coding. It can code 100 times faster than a person. And this is what
AI should be applied in crypto. Therefore therefore we don't like to be called
defy we call we leave ai we call we are just ai for crypto or ai for for defy and we are utilizing
agents to build the strategies however the strategies are you know as they were 10 years
ago they are totally verifiable end-to-end code, but they are, instead of being coded in three months,
they are being coded in 15 minutes.
Yeah.
And we are using the code on DeFi.
So yeah, yes, we do have agents,
but we don't use them personally on DeFi.
We use them to use DeFi faster and better.
Yeah.
Thanks a lot, Neo.
I gotta be honest, personally, I hate the term DeFi.. I think it's like a little bit weird and difficult to pronounce.
That's the reason we're trying to, like, you know, diversify ourselves from that.
Yeah, that makes a lot of sense. And if we can agree on another convention, I would be super happy to use it.
Why do you think we need AI in DeFi?
Yeah, I think it's important to be careful about the terminology we throw at this because
agents can have an array of intelligence and we don't necessarily have great implementations of AI as in the popular knowledge of like
LMs and LM records in finance because in order to come up with small numbers and predicting
small numbers, we need much more specialized models than generalized models.
So I think it's important to, we always like to use like agents term rather than AI agents,
because our agents as an array of intelligence from like very simple deep health models to smaller machine learning models to larger machine learning models.
But the AI part is most of the time used at the coordination layer.
I think we do need agents in DeFi.
We focus on actually automating decision making and not just automating transactions.
And I think for that, one question that the audience should ask themselves is,
do I want a mediocre life and mediocre finances?
If the answer is no, you need a better entity to manage your time and wealth, which are the two most important currencies of life.
Currently, simply put, the market has exceeded what humans are cognitively capable of digesting and reacting to.
And the financial market and human decision making has never been a good fit.
And it continues to be a worse and worse fit by day.
As humans, we are terrible at being profitable.
We are emotional.
We need to sleep.
We need to eat.
Market runs 24-7,
and we cannot keep up with thousands of rules
and protocols and chains and risk services
shifting every block.
No human can parse and forecast and act fast enough for it.
And that's where the agents come into place.
They take a high-level user intent that removes the burden of technical knowledge of accomplishing
financing goals, and they digest market complexity.
They understand all of the real-time APIs, the volatility, the LPs that most humans are
not capable of, and they
continuously execute optimized user-aligned strategies.
And one of the other reasons why I do think we need intelligence and agents and more sophisticated
decision-making in DeFi is because the intelligence of a market very much depends on the interpretive
capacity of its participants.
When users can't keep up, they revert to heuristics,
like FOMO, copying others, zero-sum games,
freezing entirely, and that creates,
kind of like more or less where we are in DeFi right now,
where capital sits idle or gets misallocated.
So that's my take on it.
I really love that answer, Aranch.
And yeah, I can definitely relate. It took me years to realize
how terrible of a trader I was. And overall now, I think I'm wiser and I prefer more automated
strategies. But I'm curious, actually. One of the things that you all mentioned is basically the automation of the unique, I would say like the unique value prop as opposed
to like a standard like yield optimization platform that may tap into like whitelisted
vaults and then auto reallocates and then things like that, which also takes the manual
quote unquote labor out of the human hands into automated strategy uh since rent you touched on this last
uh why don't you kick us off uh for this question yeah for sure i think it's it's a really really
good uh question this differentiation between like vault and script and agents and where they
all fit in place um there's a couple of couple of differentiations right first of all
like scripts they most of the time react and we build agents to anticipate so in this financial
ecosystem you can get to choose to be reactive or correct um building agents or you know building
instruments that can enable more than one layer of intelligence allows us to be
the former one where for example we don't just ape into any single apr that comes up but there's
the sophisticated forecasting mechanism that's behind it whether that's through the internal
machine learning model that we have built that allows us to forecast the apr of a given pool
for the next seven days with a high success rate,
or it is towards doing other things.
And one of the parts that we take pride on is this differentiation between vaults and GISA agents,
because vaults pool and we personalize.
Vaults assume a one-size-fits-all model where, you know,
in traditional finance, if you go to the bank
and get the simplest and the stupidest instrument,
which is a loan, it comes in a tailored and personalized manner
where we have never tapped into personalized finance in DeFi
because it was too complex.
And it was much easier to pull everyone's capital in the same vault
and move it at the same time with the same rules
when everybody has different capital requirements and risk requirements and security parameters so we take a completely
opposite approach where every single person and every single entity gets their own agent that
executes independently based on user's risk preference liquidity constraints and strategy
goals and this level of individualization historically unavailable in DeFi, is a core architectural principle.
And it's also like a passion that we have that extends beyond DeFi inventory where there is a global trend for hyper-personalized agents that run your life for you, that complete your life for you.
And we think finance is not going to be left behind.
And we like to build for that present
and for that future.
That makes a lot of sense.
The personalization is definitely something
that's interesting to explore with AI agents.
Actually, I believe this is something, Alessandro,
that Brahma is also focusing on.
I believe with Brahma Morpho agents,
you can actually select creators, Gauntlet for instance,
as a preference for the Vault optimization.
Alessandro, can you talk to us a little bit more
about the special things about Brahma and Brahma Agents?
Yeah, I think the question was between, for example, agents and algorithms first and
algorithmic optimization or bots, if you will. So on that, I think the points are either that
you're letting agents coordinate the next set of outcomes based on the output from a step. So what Neo mentioned, that the agent is actually not acting in the reasoning,
but maybe the coordination of the steps.
So let's say that you're about to rebalance the user into a certain vault
that has a new collateral or an outlier in liquidity or yield,
and you want to check in with the user before you perform
the next set of steps and checks.
So the agent will actually be coordinating the step of reasoning,
coming up with opportunity, conveying it to the user,
the user would accept or deny.
So that's a good example of a difference versus a bot, for example.
And the other thing is in our personal agent, for example, for Morpho,
there are simple and harder ways
to evaluate the best opportunity.
In general, agents are able to ingest and perform
really real-time decisions.
Generally speaking, you might have an output from a bot
or an agent to enter a new vault
that was just created recently on Morpho.
It has a new set of collateral tokens and new
stables that wasn't there in previous vaults. And so again, you can have a set of rules,
but also a set of data that can be pulled to assess if it's a good idea for that extra APY,
that much liquidity to actually enter and take risk into a new collateral. If this was a bot,
into a new collateral. If this was a bot, obviously, this new vault would have to be audited or
basically exemplified by the team into the script before it could even be considered.
So with Azagrama, one of these examples is every time an agent thinks and says,
I want to enter this new vault and reallocate you based on your individual preferences,
you have to make sure that you won't dilute the APY
so much with the size that you're entering with.
You have to check the slippage
and you have to consider multiple data points
before you take that position.
So I think the personalization is good,
but you could potentially bake it into a bot as well.
But when it comes to either coordinating
some uncertain steps or reasoning
with available data, but some reasoning capability
on top of it, such as the risk of a collateral you get exposed
to, that's when agents actually make a lot of sense
in this case.
Awesome.
Thanks a lot for the insights, Ale.
Nilo, what about Almanac?
How does it differentiate?
Yes, so we have slightly
different approach here so i'm gonna i'm gonna go through the story so um from we started in
very similar assumptions as guys have been explaining we believe that you know agents can
move the market can you know be personalized for everyone can be connected directly to blockchains
but we we face a lot of challenges
while speaking with big allocators we from back in the days like from at the beginning we've been
working with with asset managers that manage billions of dollars and most of them were very
concerned about giving any control to uh to ai there's indirect problem injection problems there
is also a famous case when agent was deliberately told
not to send money to anyone.
And eventually people created a prompt
that convinced this agent to send the money
and the agent sent $50,000.
So we knew like if you have a honeypot problem
and the agent is LLM, the black box has any sort of control
on a wallet, people will find a way. Like if the pot is big enough sort of control on a wallet,
people will find a way.
Like if the pod is big enough, they will find a way.
So we took a different approach.
We took an approach where we know that blockchain
and DeFi is great at storing value.
So we are using current infrastructure.
On this infrastructure, we are deploying a strategy
that is fully verifiable.
And we are using a swarm of AI agents
to build the strategy, to look for the strategy,
to seek for alpha to optimize it.
So we have currently 18 agents in Almanac.
They are divided into three different teams.
Tomorrow we are launching the first team,
which is a strategy building team.
And the strategy building team is a swarm of agents, which is a strategist, the coder,
the debugger, the QA engineer.
This code is also, another agent is also setting up permissions for these agents and
eventually deploy the strategy.
In this code, so this is how we are using agents.
We are using agents in front of the AI.
You communicate with these agents using natural language so it's a vibe coding file and those agents collaborate together and for example the
debugger can't find something within the strategy that is going to ask you hey the strategy is
unclear please clarify while the strategy is being created and finished then another team
is going to optimize the strategy and for example we're going to to optimize the strategy. And for example, we're going to see
if the strategy can be liquidated
or the strategy is too high.
And you can maybe change some venues.
This team is still in production.
We are using it internally, but it takes some time.
And the hardest team to find, to build, which is also
like a quarantine test, is the alpha seeking team.
So those agents are basically reporting to you, hey,
I think this strategy is going to gonna be great we should code it but for the for the for the thing i mentioned
at the beginning in order for big money to deposit into this agents the strategy has to be end-to-end
it can't evolve unless you allow this strategy to evolve so our workflow is you build the strategy using the AI you seek alpha using
the AI but one of the strategies while the new alpha is being found while the strategy is created
and optimized um you you are making the call whether to deploy the strategy you are running
a back test you see the strategy works but this is very similar to the traditional hedge fund
workflows and we got
the feedback and we already see that it's working, that big money is interested in that
because basically they can fire all of their quants team and they can decrease the time
of deployment of the quant strategy from months to minutes. This is our approach. It's way
different. The strategy is very key deployed as an end-to-end code onto the blockchain,
into the vault. You can create it first privately on a safe wallet, so all the permissions are under your control. So also the vault owner or the strategy owner can control whether
the strategy can't or can do with that. And if the agents find a new opportunity or how to update
the strategy, they are free to do so, but you need to confirm it opportunity or how to update the strategy they are free to do so but
you need to confirm it you need to update the permissions so we have full control it's like
block you have full control through the blockchains you can use ledger you can use
treasure and so on we have slightly different approach here i'm not sure which which is better
we you know we are targeting big money and yeah it's so far so good uh we got uh one of our team just deployed the
first strategy like a week ago we are working with more teams and they are deploying more strategies
uh so yeah awesome thanks a lot uh you prefer for all the insights i think it's very interesting to
see how you're each tackling uh this particular i was gonna call it an issue tackling this particular, I was going to call it an issue, but this particular
aspects of yield earning. I'm actually curious because I think all three projects have been
around for not going to call it a long time because this is all pretty novel space, but
at least they've acquired like a certain amount of TDL, have grown significantly in the last
few weeks. And I'm actually curious if any of you have any kind of like, you know,
metrics or case studies or insights into particular examples in which your
model was successful in, you know,
capturing a novel yield source or generating better APY than classic vault
or like any kind of insights like that.
Neil, because you wrap this up, you're on the roll.
Why don't you kick us off with this one?
Sure. Yeah. So it's great.
We just, like I said, one of our core contributors launched the first pool.
And at the beginning, the pool was pretty decent let's say the strategy was fully designed
by AI and the strategy was very simple was just seeking the highest yield across Morphe,
Euler, Yern, compound I think and fluid and the strategy was basically just depositing
between USDT, USDC and I think that, yeah.
Very simple strategy, but as guys mentioned before,
you can sleep well, you don't have to rebalance every day.
The beauty of it is that the strategy was designed by AI
in like a couple of hours because it was back in the days
when our AI was still a bit slower.
Now you can design strategy like in a couple of minutes,
hopefully.
However, we got the feedback from, again, big allocators
that the strategy should be slightly less degen,
so we decreased the risk profile.
But I know that the guys are already working on it,
working on making more strategies with more,
more, let's say, degeneral risk profiles.
So yeah, I don't have much to ask here. It's pretty simple. with more, let's say, degenerate profiles.
So yeah, I don't have much to add here. It's pretty simple.
The strategy itself looks for the highest yield
and optimized and allocator.
The beauty is that it has been designed by AI
and the AI can always optimize the strategy
if the strategy owner allows it.
It's very cool, it's wrapped in the vault.
So it's fully fungible, fully composable.
It can participate in further DeFi legal.
I know that the guys are already working on adding
it to Euler or to other lending protocols.
So yeah, the best from the vault world.
Very far scouting from the AI space
and full composability with DeFi on the blockchain space. I love the composability angle. That's really cool.
Rensch, any insights on like ARMA agents performing or Giza agents?
I'm actually not too sure how you would call them because you called them Giza agents,
but I understand that Giza is the infrared layer and ARMA agents are actually the name of. First thing
you please clarify for us exactly. And then two, do you have any insights, any interesting
metrics or any interesting examples in which you feel like your products outperformed traditional
gear sources?
Of course, of course. The first to verify is is infrastructure that allows for autonomous finance to be built on top.
Arm is the first agent built on top of Giza because as an infrastructure project,
you don't really want to be another infrastructure project in the tree that gets nothing built on top.
And I think as an infrastructure project, the best way you can differentiate yourself is to build a product on top of it
that has a great use case and that has product market fit
and that's why we built armor on for ourselves and army has been live uh since five months so
we do have a lot of data and we do have a lot of back testing i'm happy to share those here um armor
let's see um the base yields on base for stablecoins is around 4%.
This is the average of any static strategy.
And Armour gets you two days since launch around 8.6.
So, with the capital optimization, we're also taking into account
what I mentioned about, like forecasting cost, understanding where the…
Right, and Shabzai I'm sorry just sorry to interrupt
you your mics sometimes gets a little bit far and it's difficult to hear you uh just a heads up uh
if you're able to do anything about it sounds good can you hear me now yeah and it's much better
perfect uh so as I mentioned um army is averaging around 8.6R on stables since its launch on base,
while the average of static positions is around 4%.
And through continuous capital optimization,
we're able to place the users at the most profitable position at any given time.
And this takes into account the risk-adjusted yields,
so we don't ape into any APR that comes to surface.
And I do have a great story to tell
with the numbers because we do have a product out there um it's happened actually last week
that kind of shows uh what we're about um so between 27th of july and first of august this is
like this five day period arma has moved $230 million in volume.
That's like around $45 million per day and around $2 million per hour and $30,000 per minute.
And in this span, we have executed around 43,000 transactions that were 100% profitable
for the thousands of users that we have that require zero signatures.
And most of it happened while people were sleeping and that is towards like the wealth part of performance
but also there's very much the time part of performance we have this internal kpi called
return on attention which is not just return on investment but the attention and the quality of
which is not just return on investment,
but the attention and the quality of life,
where these agents that allow us
to automate decision-making on behalf of users,
they do not just get you money
through intelligent automation,
through information distillation,
through simplified interactions,
through cognitive ergonomics.
For users, we unlock complex strategies,
but we also reduce burnouts we also
save invaluable amount of time and we democratize you know participation to the
and that's like another measurement that we like to take when it comes to performance so
for sure we'll get you to x more yield but we will get you multiple times of of your time as well
awesome thanks for the insight range that's a that's a lot of volume uh for that's a lot of multiple times of your time as well. Awesome.
Thanks for the insight, Rancho.
That's a lot of volume.
That's a lot of agent volume.
Alessandro, what about Brahma?
So we launched the agent around January this year.
And I guess a lot depends on the number one,
the rebalancing frequency we've seen from the users.
And also another thing we look at is how do the users set up their agent when they start?
As you said, they can customize minimum liquidity, minimum API, Curator, for example.
And so every agent acts totally differently based on the constraint the user gives it.
What we saw is that on average, there's up to 30% of performance from the agent versus a static allocation.
But that really depends on the rebalancing frequency because you can come in.
We have a lot of users that come in with a very small sum of money, and then we have users that come in with 500k plus tickets. Obviously, the gas
and the rebalancing frequency of the agent then changes because it doesn't want to also eat into
the yield with the gas feed and so it's not always rebalancing in the same way. What we noticed,
which was counter-intuitive to what we thought, because when we set the liquidity filter, it already has a liquidity filter by itself, like a threshold.
So let's say you deposit 50k,
it will only consider vaults when rebalancing
that have an average liquidity of above 2 or 3x,
depending on the user profile.
So it doesn't deposit somewhere where it's too small
for the user position.
But then the user can also set their threshold, as we said manually.
So we saw that most of the users actually set a very tight threshold on liquidity because
they're looking for a riskier yield.
So like what Neil said on institutionals or bigger user definitely applies, but we saw
that from more DGEN medium-sized users with an average ticket of 25 to 50K,
they tend to be pretty aggressive.
They tend to select minimal liquidity of 100K.
So they really are looking for constant rebalancing
and cost-constant yield allocation.
So yeah, these are, I guess, the main thing
we saw, which weren't obvious in hindsight.
Thanks, Saul.
This conversation about reallocation
and rebalancing frequency is actually a perfect setup
for my next question, which is that I've noticed that, first
of all, Brahma has a Morpho base agent,
and Arma agents are available on base, actually on base app.
Neil, I know that Almanac, while the only vault available currently is on EVE, is actually
compatible with base, like overall, like your infra is compatible with base.
And then most of the composition, the current composition of the EVault is on Morpho. It's very interesting that Morpho and Base come back so often into current AI agents' projects.
And so I'm really curious about why that choice.
And I can take a guess, gas fees, permissionless deposits, all that kind of stuff.
But I would love to hear it actually from you and starting with Ranch actually.
How, like, why the choice of base and Morpho particularly as like major allocations or actually infra to build on and chains to build on?
Yeah, there are obvious factors which we consider when picking ecosystems.
And L-ploofs are a natural fit given the low gas environments they provide.
Although this is not a solved problem since what agents desire is the
zero marginal cost of exchange so they can act with utmost velocity and rebalance near real time
but base is one of the best places for this second is the multi-dimensional quality of their defy
ecosystem like the variety liquidity security incentives they all play a major role here
and lastly alignment with the partner
makes a world of difference.
Are they willing to support up and coming ecosystems?
Are they genuinely excited to bring these capabilities
like accessibility to their end users?
Do they organically hold that interest and priority high
and demonstrate that all plays a role?
So the technical parts are all there for base, the ecosystem is all there for base.
In terms of Morpho, at its core, our agents has an on-chain decision engine that ranks every volt by risk-adjusted yield per gas.
And the agent continuously solves the complex problem of earning more versus spending on fees and this
end takes like a smart modeling forecast deal for every vault from live market data it estimates gas
across l1s and l2s to know exact cost of moving and it also checks guardrails that are set by
users and morfo in this equation keeps on winning highest Highest risk adjusted yield per gas and deep liquidity.
And a great partner. Awesome. That's good to hear. And yeah,
I really like the Arma agent interface on Basehub. I think it's really cool.
Next, Neo, what about you? I know your main vault right now is live on EVE. And so I'm curious about the expansion to other blockchains
because on Almanac's website,
you can see all the integration partners,
all the integrated chains and all the integrated protocols.
And I'm really curious about your plans for expansion
and kind of like where you see the puck going.
Yeah, sure.
So internally we currently support almost any VM, and you can even build strategies on Texas, centralized exchanges using our swarm of AI. However, they are not connected yet. Tomorrow we are releasing the integration with Ethereum, arbitrum and base so you would be able to deploy a strategy on
this two on this three um the reason why very similar to what ranch said uh arbitrum and base
has great gas fees tomorrow we are not we will not allow users to deploy volts um in a permissionless
manner yet it's likely October so base is great if you want to deploy vaults in a permissionless manner yet.
It's likely October.
So Bayes is great if you want to deploy strategy for yourself,
deposit like 100 or 1,000 bucks,
then the gas cost is not going to consume you, yeah?
We're likely going to be enrolling more and more chains.
Obviously, Hyperliquid is on the roadmap. Currently, I think we support eight chains and 150 protocols
that you can Vibe code on.
Yeah, when it comes to integration with Texas,
likely probably end of the year,
maybe beginning of the next year,
probably my team gonna kill me for those days,
but let's see, let's hope not.
And when it comes to choosing the ethereum it wasn't actually
our choice it was the choice of of of the contributors that have built the strategy and
it was up to them where they want to deploy the vault and they said okay ethereum still
has the biggest assets it you'll say still have the biggest players and if you want to build like
a billion dollar vault what they are trying to build it's still the ethereum uh I'm not sure where they are going to deploy the next vault.
They've been thinking about either base or Binance even,
because the vault is going to be more risky,
but it's eventually up to them when they're going to see
the biggest upside when it comes to allocators.
Thanks, Neo.
I love that Neo is just announcing a bunch of stuff on the spaces.
I was actually, I was like, oh, I was surprised. Like I was checking like Almanac's Twitter.
I was like, oh, none of this is actually out. So making like the alpha and the big commitments
here, I really like it. Ali, Brahma is a little bit particular and specific because you actually have a specific base Morpho Agent
that is out.
Why a base Morpho Agent?
Yeah, actually, the base Morpho Agent was the first,
but not the only Agent we built.
And because we released a developer
kit called ConsoleKit
that runs on any EVM chain, and now we're expanding it to non-EVM,
we actually have developers that are building on any other chain.
So for example, right after the Morph agent, we also built...
Actually, it's a T-Watt, like a swap agent that has some reasoning
in terms of order sizing and binding when you want to either
accumulate or sell a token.
And that's on base as well, simply because we have like bases really fast, like fast blocks,
sick fees. And so we have a really good execution system there. But last month, we also expanded to
HyperVM with Felix, the CDP, and we built an agent with them for actually managing your true interest rate
autonomously.
The reason why we launched Morpho and Base, well Base Performance of the Chain and Morpho
initially, was because it actually has an interesting amount of manual work required
for the users to be able to basically continuously or
pretty frequently evaluate vaults, move their position around. That wasn't there with other
protocols that have kind of pooled lending markets. And so with Morpho and the proliferation of vaults,
it felt like it makes a lot of sense. And now we're expanding this to this and other logic to any other chain.
But again, it's the developers building on it.
And we kind of build this just as use cases
to showcase and scale on specific protocols that
are aligned, as the guys said.
So yeah.
That makes a lot of sense.
Thanks all for your insights on this.
Actually, let's switch gears completely.
I think I want to talk about something that each of you touched on at some point earlier during the space, which is the risk management and, like, the safety and the wall rails, right?
So this all sounds great.
Like, you know, personalized preference yield, swarms, different vaults, different kind of, like, ways to earn yield and, like, swarms, different vaults, different kind of like ways to earn yield and like swarms
seeking different alpha.
Like this all sounds great.
But what about like some of the risks that come with it and how do you mitigate them?
Like for instance, you know, how do you prevent like AI from like chasing just pure yield
maxing at the cost of excessive risk?
How do you handle like the additional like smart contract risk on top of it?
Market changes, like what kind of like an event of like, say, like a DPEG,
like all those things could very well happen, right?
How do you actually handle those potential events or what kind of ground rails do you put in place?
Neil, why don't you start with this one?
Awesome. I love this topic.
So, like I said, we are not reinventing the wheel.
We are using what is already working
so we basically copied a workflow of a traditional hedge fund yeah so um we are having three teams of
agents um one of the team let's say let's make a use case let's say a yield second agents or the
team of agents that looks look for alpha, found some alpha, and they're
proposing this alpha. Let's say let's build a funding rate
arbitrage between Binance and Hyperliquid. Yeah, agents found
it and they're proposing it, hey, you can make I don't know 20%
a year on that dollar. We are saying oh, it is great. Let's
code it. So then you are transferring it to a different set of agents different group of agents a strategy coding agents those strategy
coding agents gonna code it for you and you have fully verifiable deterministic and end-to-end code
this code is finished it will not change it will nothing gonna nothing bad gonna happen with this
code so now what you can do with this code,
you can backtest it, you can optimize it,
you can simulate it.
So you can do everything what a traditional hedge fund
would do with the code before deploying it on chain.
And here comes the third team of the agents,
which is the optimization.
So this team of the agents optimizes this code
basically by trying to find for this particular strategy
optimal venues venues optimal risk
reward ratios optimal sizes optimal assets and so on yeah so we are outsourcing the process of
building the strategy and optimizing the strategy to the ai and once the strategy is finished
it's it's finished it's uh it's up to you now to sell the strategy if you are a you know a person
who created it together with the agents you are unlikely going to share the to sell the strategy if you are a you know a person who created it together with
the agents you are unlikely going to share the code of the strategy because then you are giving
up all the alpha and all the capacity profit sharing and management here are gone because
you you you gave up your your strategy likely gonna create a vote yeah you're gonna accept
capital into this vote and it's up to you for example, create a risk report or to find a third party auditor that's going to audit this code and you're going to pay for it.
Or you're going to find a third party, maybe insurance provider or something like that to make people more comfortable to deposit into your vault.
Yeah, so we are leaving it for the market.
Market is going to figure out how to basically make people comfortable
investing into strategies created by our AI.
But we are not reinventing the wheel.
We are not applying some God knows what.
It's the same workflow as in traditional hedge funds.
They've been around for years.
People are investing billions of dollars into hedge funds, quant funds.
And yeah, we are just using AI to make it faster.
Makes a lot of sense. Thanks, Neo. Ali, what about... Maybe I can add on the blockchain side. Sorry for that.
Because of the blockchain side, the strategy has permissions.
So strategy is connected with the blockchain using the permissions.
Permissions are fully auditable, fully transparent, fully verifiable.
Anyone can see what permissions the
strategy has so the strategy can't do anything funky with your money it can for example only
deposit from other withdraw from other deposit into other withdrawals for example yeah so those
permissions are you know fully auditable on chain minted nothing gonna happen yeah yeah that makes
sense uh that was the other part uh which is like, yeah, of course,
the smart contract is going to be all in-chain risk.
Ali, what about Brahma?
Bit of a different approach, but as a side note,
when we started building the info for Brahma,
that was initially like B2B smart accounts with permission
settings.
So essentially, we were working with fund managers and treasuries,
which are still some of our larger clients.
And we had to build a sub-account system that had a policy layer on top
that you can give instructions to.
And initially we started thinking of all the edge cases that a manual operator
could try to do to essentially evade controls and withdraw some fund.
This was the idea of if you want to give access to a specific person to
trade or manage only portion of your capital in a subaccount.
So we built a whole policy manager out, which is a bit similar to what a custodian would have,
something like Fireblocks or Bitco, they have their policy management systems. They run off-chain in their case, and in our case, it's a mix of off-chain
plus on-chain guards that are applied to the user's smart accounts.
And so when agents started coming up, or in general, we built this for first
manual management and then bots, we started having people spreading through API
on this sub-account, and essentially, you just have to scale the conditions and the constraints that
are put off-chain.
And they can be very expressive.
They can be time constraints.
They can be caps, a whitelisting of contracts of tokens,
input output, and other non-deterministic edge cases,
such as when you're executing a big multicoll
that wouldn't even be possible to be fully decoded on-chain.
So essentially, that was one of the things we saw when agents came up,
and you have this issue of hallucination.
Whenever an agent or a manager or a bot wants to invoke a transaction on a Brahma user account,
they have to, first of all, to ping the policy system that will essentially
pre-simulate and post-simulate this. Even the output from what the user signed as the policy
with their key on their account, if it goes through, then it goes straight to the execution
system with the relayers and the RPC balancers that will execute the transaction. So essentially, this is a very strict policy system we built already before we started
this in 2023.
And then we had to adapt it to a bit of the input conditions that an agent would have.
But essentially, it's a very flexible system.
And the rules stay private because the clients initially didn't want to show their permissions
that they set individually on-chain. So that's why the hybrid model. because the clients initially didn't want to show their permissions
that they set individually on-chain.
So that's why the hybrid model.
Yeah.
Thanks, Ola. That was super interesting.
Ransh, why don't you wrap up this question?
Yeah, obviously one of the biggest questions.
Most of the time, I think, especially the public has has this scare of will the agents run away with my
money or will ai as i mentioned in the beginning of the call like the general understanding of ai
with alams generative models will hallucinate into like a different financial past um starting there
we use fully deterministic and explainable and traceable and auditable and verifiable models
and algorithms for our decision making and we have a non-custodial safe and policy-bound
automation environment so every user gets an account abstraction module that acts as an
extension of the eua with their with after this module once they deposit their capital
they sign basically a contract between the agent and themselves defining the parameters the agent
is going to operate within and once those are assigned those are assigned cryptographically
and the agent cannot operate outside of those boundaries so these are quite critical components
that happens on an individual layer.
There's also a global level where we continuously monitor health can withdraw funds where liquidity allows and put in the right
emergency protocol where where it dictates so we do not only automate decision making but we also
automate risk management awesome thanks ranch i love how how it's very clear that each of you has thought very,
very deeply about this and how to actually counter for the specific edge cases of your
individual infra. It's really clear it's deeply thought through and not like an afterthought
in terms of just chasing the yield, just making, just chasing the yield
or just making like the best strategy out there.
But it's like, you know, clearly having the AI layer actually opens up to like a variety
of different risks and age cases and it's all been very thought through.
So this is really insightful and thanks all for your answers. It's super interesting to see how you're tackling individually all those potential risks and cases.
All right, so we've got one more question actually.
And this time we're kind of like looking ahead, which is how do you see AI-driven yield optimization evolving?
I know next few years seems like a lifetime horizon in terms of AI,
but at least in the next coming few months to a year,
do you reckon it will be something that becomes mainstream?
Do you think it will still be for power users
with deep knowledge into DeFi and specific preferences
and very comfortable with their
risk profiles.
What kind of improvements do you see on the horizon?
Is it extending to any kind of yield sources such as RWAs and things like that?
I'm just keen to know where do you see the wider adoption of this?
Because obviously all three of you building into the space are convinced uh hopefully
that uh you know there will be growing adoption of these kinds of products i'm just keen to know
like where do you see this adoption happening and what are the um the evolution that will drive uh
this adoption and for this last question um let's uh sorry, Ale, why don't you kick us off with this one?
I think our opinion is it will take time.
And I think the guys like Neo even mentioned,
like, institutionals or bigger depositors
are skeptical of agents.
I'm thinking more from a distribution standpoint.
Like, there's a lot of new interfaces and crypto yields or crypto credits
entering fintech through mostly vault structures for now.
And even in these vault structures, the curators today are still mostly,
and the risk managers are still mostly executing everything manually.
So when I look at that and we talk to them, for example,
of general automated
execution, that is even not the common norm right now. So thinking that systems that scale and get
distributed to not-crypto-native users will use AI out of the box doesn't feel like it can happen
right away. But essentially, that's how I think it will happen. It will be enshrined into products that are offering you deals.
They have deeper or lighter levels of AI optimization built on top for something.
It can be risk management.
It can be for maintaining the health factor of a loan with some consideration on top.
It can be personalizing the product for the user or extending it to them.
I think that's how it's going to be. It's going to be packaged inside a product that already
has institutional stamp of approval. Let's call it the Veta Vault or something else.
And then I think there will be some teams that bring direct-to-consumer applications
that are offered already with this logic built in. I just think it would take a little bit of time.
But the good thing, I think Rens said it,
that we're in a programmable system,
and therefore agents can talk to each other,
they can subscribe to each other, and cross-agent collaboration
can come much faster than on traditional Rails.
So that's the good thing.
I love that endnote. Rensh.
Yeah, I think in our stance, we expect a massive adoption due to a radical lowering of entry barriers through solving all of the issues that we discussed on this call.
through solving all of the issues that we discussed on this call um there's going to be a huge
you know new wave of users and capital coming into d5 you know and this this idea of this like
memetic objective has been thrown around in the past decade in web3 by almost every project i
think oh like we are onboarding next billion users to web3 to DeFi well you know the
next billion users are not going to come and use the current DeFi products they cannot it's it's
in inhumane products that require excessive amount of technical knowledge to conduct and
in this trend of hyper-personalized agents where applications are disappearing where you know additional layers
are disappearing the same will happen to defy and and is already taking place um the protocols the
vault the chains they will all become a back-end component and the main conductor of capital for
the users and institutions alike are going to be the agent interfaces and this will constitute
a total transformation of how we define liquidity in d5 and and for us that goes through the
evolution of this metric called tvl which which which is you know quite quite a stone-edge metric
into active liquidity once money moves according to fundamentals-driven analysis
at near-zero marginal cost, capital will be instantly available
for any products which provides the necessary utility.
Trust management and incentives, moving away from static zombie pools
and protocols, increasing the pace of innovation
and massively improving capital
efficiency is going to change what d5 is what liquidity is in d5 and who can conduct it
and that is you know super exciting i love that i know i also think it's super exciting um
neil what do you think awesome yeah so uh again i have a little bit different opinion um i don't think
we're gonna ever see big big money depositing money into the to be managed by ai i think ai
uh gonna dramatically increase how we manage like increase the efficiency of how we manage money so
basically discovering alpha building new strategies and so on but overall it's still going to be deterministic codes uh unless the you know someone can verify that this llm is not going to do
something funky with my money no one is going to put more than 100 bucks so we are thinking a little
bit different with that we are thinking more about scalability of money how much money can be enhanced
with ai management and this is this is our target basically we are targeting as much money can be enhanced with AI management. And this is our target.
Basically, we are targeting as much money and capital
that can be accommodated within the infrastructure
that is enhanced with AI.
So this is our approach.
When it comes to retail, I think,
regardless of what's going on behind the,
like under the curtain or like on the backend,
it doesn't care if it's AI or not,
it just can't lose money.
Whether it is a vault, whether it is a wallet
or something else, we still gonna have to provide them
probably a very simple app
when they can just ape in on the app from the phones.
And this is not our target.
If someone wants to build a app like that on on the top of
almanac infrastructure we are happy to you know to accommodate that uh but yeah i think still there
will be a bunch of people that are smart enough to really leverage ai and you know squeeze the market
more uh but i am a little bit skeptical about every single person having a personal agent within their
their phone and telling them i want to save for for a car or a house or something like that i don't think it's
gonna ever happen first of all people are you know tick tock brains they won't be even able to fill
up the form and second of all uh once the ai gonna be exploited or gonna lose the money the business
is that yeah no one's gonna deposit money there even and it's not uh if gonna happen it's when gonna happen the more money is being managed by the honeypot the
more people gonna try to exploit it i highly encourage everyone to see the the uh to check
the in the article about uh the wallet which was deliberately told not to send money to anyone
every single uh prompt was costed i I think, a couple of dollars.
Eventually it was $50,000 and someone managed to convince agents to send him the money.
This is our opinion.
We are scaling liquidity sites, not the number of users sites, and the users will come on
the top if someone wants to.
If the reward is high enough, let's say like that thanks neo that makes
a lot of sense i actually played that jailbreak game uh well i'm not i don't even remember i think
it was uh uh it was freya uh that was a lot of fun um and yeah it will cost like increasingly
amounts of if uh if you were able to actually jailbreak the AI
and get it to send you the treasure.
The prompt is super interesting.
There is a whole, actually, a little side note, there is a whole interesting universe
in terms of prompt engineering for those kinds of jailbreaks that are super fascinating.
So, if you are curious about the topic, there is a lot that has been written on it.
You can do some searches.
We're almost at time.
I'm sorry we're not going to be able to take any questions
because I'm conscious of our speaker's time.
But I promise you that if you thread a question
on Gauntlet's Twitter when we post this,
when this space is over.
I will chase whoever's got the best answer,
and I will get you that answer.
I make that commitment here live.
I'm happy to stay like one hour after
and answer all the questions.
One hour.
Unfortunately, like, I can't stay.
I myself can't stay.
I'm like in the comments, yeah?
Like, for the next hour,
I will be answering the questions in the comments, you guys.
All right.
That sounds great.
That sounds great.
Let's do that.
And now any final,
Nio, you already made six announcements on this basis,
but any kind of like announcement, insights,
any parting thoughts that any of you would like to make?
Now is the moment.
I'll give you a minute or two if you want to.
If no final thoughts, that's also absolutely fine.
I can maybe focus on one thing.
No, I think you're good.
Go ahead, Nero.
Go ahead.
So, yeah, tomorrow we are launching our AI forum to build strategies.
So stay tuned.
It's, I think, 10 a.m. UTC.
It's going to be first 100 people.
Yeah, that's it.
Would love to have you on board, guys.
Thanks, Neo.
Ryan, she had something?
No, I was just going to say thank you, everyone, for joining and going into this intersection.
The next few weeks are going to be pretty packed for Giza. The new launches coming. So keep an eye out and appreciate everyone this intersection. The next few weeks are going to be pretty packed for these new launches coming.
So keep an eye out and appreciate everyone's time.
Thanks, Rang.
Ale, any kind of final thoughts?
Yeah, thanks for having us.
I think we need more teams and more devs
knocking their head on this until we find the right setup.
So good to get everybody together. Thank you.
Thank you. I really enjoyed having you free on the space. The discussion was awesome.
I really, really, you know, really enjoyed that conversation. Thanks to everyone who
tuned in. This was great. Stay tuned for more AI agent content from those amazing speakers, their companies.
Follow Almanac, follow Brahma, and give a follow to Giza as well on Twitter and, of course, Gauntlet.
And, yeah, stay tuned for more AI agent content as we dive deeper into the frontier of DeFi.
Thanks, everyone.
I hope you all have a great day or night, depending on where you are in the world.
Bye. Thank you, everybody. Thank you.
