DeFAI is the combination of decentralized finance and artificial intelligence: software agents that read onchain and market data, decide on a financial action, and in many designs submit the transaction themselves. Chainlink's education hub, last updated February 11, 2026, defines DeFAI (also written DeFiAI) as autonomous AI agents that use blockchain protocols to execute financial strategies, optimize yield and manage risk (Chainlink: What Is DeFAI). This page separates the two kinds of DeFAI agent that most explainers merge, agents that advise and agents that sign, and maps where execution risk sits in each.
​

Last reviewed: Nov 2026. Every figure below links to its primary source in the same paragraph.
​

DeFAI means AI agents acting on DeFi protocols. The agent gathers data such as prices, balances and pool states, a model turns that data into a decision, and the decision becomes a swap, a deposit, a loan adjustment or a transfer. The term covers both research tools that only recommend and autonomous agents that hold keys.
​

The difference from a classic trading bot is the decision layer. A bot follows fixed rules written in advance. A DeFAI agent uses a language model or another machine-learning model to interpret goals stated in plain words and to pick among actions. Chainlink describes the shift as moving from static "if-then" smart-contract logic to agents that monitor conditions and act continuously (Chainlink).
​

Eco's knowledge base already covers the adjacent categories: agentic trading, market-intelligence agents such as AIXBT, and agent frameworks such as ElizaOS. DeFAI is the umbrella term these sit under.
​

A DeFAI agent runs a loop with four steps: collect data offchain, compute a decision offchain, deliver that decision to the chain, and execute it in a transaction. Blockchains cannot run large models directly, so the reasoning happens offchain and only the final instruction, signed by a key the agent controls, reaches a smart contract.
​

Chainlink's version of the loop places an oracle network between the model and the contract to verify inputs before execution (Chainlink). Agent toolkits that connect a language model directly to a wallet skip that step: the model calls a tool, the tool builds a transaction, and a wallet signs it. The Model Context Protocol, an open-source standard for connecting AI applications to external systems, is now the common way these tools are exposed to a model (MCP documentation).
​

An advisory DeFAI agent produces analysis, alerts or a recommended trade and leaves the transaction to a person. A signing agent holds or controls a key and submits transactions on its own. The model can be identical in both; the risk profile is not, because only the signing agent can lose funds through its own mistake.
​

The distinction shows up in what agents actually buy. Turnkey's analysis of Coinbase's x402 Bazaar found that crypto market data and trading signals made up 48,853 of 208,415 calls, or 23.4%, across the top 100 listings in a 30-day window ending September 13, 2026, and noted that most of those feeds served a handful of payer addresses (Turnkey). That is the advisory half of DeFAI at work: agents paying per call for token data before a decision. Whether the decision then becomes a transaction depends on how the agent was wired.
​

The main DeFAI use cases are yield optimization across pools, dynamic risk management for lending positions, natural-language trade requests, treasury rebalancing and paid data lookups that feed a trading decision. Each use case can run in advisory mode or signing mode, and the same task carries different risk depending on which mode a team chooses.
​

Agents scan liquidity pools and move funds to the highest-yielding option after accounting for gas and slippage, as Chainlink describes it (Chainlink). In signing mode this is the riskiest use case, because a move requires a withdrawal, often a swap, and a deposit, each of which can fail on its own.
​

Lending protocols use fixed parameters. A DeFAI agent can watch a position and add collateral or repay before a liquidation threshold. The check that the position is safe after the action belongs inside the same transaction, otherwise the state can change between the check and the action.
​

A user says "put 1,000 USDC into a low-risk stablecoin yield" and the agent chooses venues and routes. Here the model is translating a goal into concrete calls, which is exactly where hallucinated addresses or wrong chains enter. See prompt injection attacks on agents that hold funds for how outside text can redirect that translation.
​

Agents increasingly buy their inputs per call over x402. Turnkey counted 15,532 onchain RPC calls and 87,567 web search calls in the same 30-day Bazaar sample (Turnkey). The x402 protocol explainer covers how those payments settle.
​

Execution risk in a signing DeFAI agent sits in three places: the model's decision, the transaction the tool builds from that decision, and the state of the chain at the moment the transaction runs. Most DeFAI writing covers the first. The second and third cause the losses that a better model cannot prevent.
​

Research benchmarks measure the decision layer. CryptoBench, an expert-curated benchmark of 50 questions per month, found a "retrieval-prediction imbalance": leading models retrieve data well but are weak at predictive analysis (CryptoBench, arXiv 2512.00417). The Agent Market Arena benchmark, which ran live crypto and stock trading agents, found that agent framework design drove more of the variation in outcomes than the underlying model did (When Agents Trade, arXiv 2510.11695).
​

The execution layer needs its own controls. A swap built from a correct decision still fails the user if the price moves before the block, if the pool chosen has thin liquidity, or if a balance assumed at decision time is gone. The standard answer is a minimum-output or balance condition that is checked during execution, so the transaction reverts instead of completing at a bad state. Simulation before signing catches some of these failures, but state can change between simulation and inclusion.
​

Sauce, Eco's onchain harness for agents, sits at the execution layer of a signing DeFAI agent. An agent writes a Sauce program that reads balances and prices and states the conditions a transaction must meet; if a condition fails during execution, the whole transaction reverts. It does not decide what to trade.
​

Eco's documentation describes a Sauce program as a set of onchain actions and the conditions they must satisfy, combined into one transaction that reverts if an unhandled failure occurs (Eco docs: Sauce introduction). Examples in the docs include comparing two swap quotes and using the pool with the higher quote, enforcing a minimum output, and checking a lending position after supplying to it. The Sauce MCP server gives a coding agent tools to compile, simulate and build those transactions, and its transaction-building tools return unsigned transactions for a wallet to sign (Eco docs: Sauce for AI agents). The CLI builds and sends transactions on Ethereum, Base, Arc and Solana.
​

DeFAI as a category has three honest limits: the term is used loosely, model quality on predictive tasks is still weak, and execution guards protect against bad states but not against bad goals. Teams should read any DeFAI claim with those limits in mind.

These are the questions most often asked about DeFAI and how DeFAI agents handle transactions.
​

Yes. Chainlink notes that DeFiAI is often styled DeFAI; both refer to AI agents operating on DeFi protocols (Chainlink).
​

Only a signing agent does. An advisory agent can work from read-only data and hand its output to a person who signs.
​

No. Onchain transactions are irreversible once included, which is why Chainlink lists model error as a direct path to permanent loss (Chainlink). Conditions checked inside the transaction prevent some bad outcomes before inclusion.
​

Key custody, spending limits, simulation, in-transaction conditions and a way to stop the agent. The AI agent security checklist and the kill switch guide cover each item.
​

Definitions and figures come from the primary sources below, fetched on November 12, 2026. Bazaar call counts are Turnkey's published analysis of Coinbase data, not figures produced for this page.