Almanak launches Agentic DeFi platform that turns prompts into live on-chain quant strategies

One prompt, one live on-chain strategy

Almanak has officially launched its Agentic DeFi strategy platform, a system that takes a plain-language trading idea, turns it into Python code, tests it and deploys it on-chain to manage real funds. According to a detailed write-up on dev.to, the launch landed on September 1, with backing from crypto investors including Delphi Labs and HashKey Capital.

The advertised workflow: a user describes an idea such as supplying ETH/USDC liquidity on Base while hedging impermanent loss, and a team of AI agents produces readable, editable strategy code, backtests it, simulates it and finally deploys it to run live. Almanak positions the product as a personal, agent-driven hedge fund — quant tooling that once lived inside institutional desks, opened up to anyone.

Eighteen agents and a heavy testing pipeline

Per the dev.to post, Almanak claims 18 specialized agents organized like a quant research department, covering ideation, coding, backtesting and risk management, rather than a single model handling everything. The post's author adds a caveat: agent count is not a quality signal. In multi-agent pipelines a small error at one stage can compound down the chain, and latency and token costs accumulate; whether the division of labor converges reliably is something only sustained real-world use can prove.

On validation, the platform stress-tests strategies across more than 10,000 Monte Carlo scenarios — checking whether a strategy is consistently profitable across many simulated market paths or merely lucky on the one timeline that actually occurred — and can dry-run a strategy on a mainnet fork before any real capital moves.

The launch covers seven chains from day one: Ethereum, Arbitrum, Base, BNB Chain, Polygon, Optimism and Avalanche. Strategy categories include liquidity provision, yield optimization, delta-neutral approaches, and tokenized stocks and commodities.

Scoped permissions, not custody

The dev.to author, who says they have built their own multi-agent trading system end to end, argues the natural-language part is the least interesting bit. Getting a model to emit professional-looking strategy code is no longer difficult; what determines whether accounts get wiped is what the agents are allowed to touch.

Almanak's answer, as described in the post: strategies deploy to the user's own Safe smart wallet, and agents receive scope-limited permissions through Zodiac Role, meaning they can act only inside protocols and action ranges the user has pre-approved. The private key stays with the user throughout. The post likens the setup to valet parking — the driver can move the car but cannot rifle through the trunk or keep it.

The author is careful to draw a boundary, though: this architecture secures permissions, not profitability. It can stop an agent from moving funds somewhere unapproved, but it cannot stop the strategy itself from being wrong, from being arbitraged, or from losing returns to slippage. Those are separate layers of risk, and one kind of safety should not be mistaken for the other.

Backtests are not returns

The 10,000-scenario testing and forked-mainnet dry runs are, in the author's assessment, sound engineering — but expectations still need tempering. The gap between a good backtest and a strategy that earns with live capital is where traders tend to get hurt, and the classic trap is overfitting: parameters tuned until they look flawless on historical data when they have really just memorized the past. The post prescribes out-of-sample testing, walk-forward validation — letting a strategy learn on the first half of the data and grading it on the second half — and general suspicion of results that look too good to be real. Almanak can help users write and test strategies quickly; judging whether a strategy will actually be profitable going forward remains on the human.

Why it matters

Almanak is a concrete case of agentic AI crossing from chat into moving real money on public infrastructure, and its safety architecture is worth studying well beyond DeFi. The dev.to post distills broadly transferable rules: give an agent the narrowest permission scope a task actually requires, because over-broad access is a when-not-if problem; force agents through consequence-free dry runs — sandboxes, forks, test modes — before wiring them to real side effects; and never confuse a well-contained agent with a correct one. As agents increasingly book, file, trade and pay on people's behalf, how small the permission box is may matter more than how smart the agent inside it is.