REI Labs
Every action an agent proposes passes through a second layer of intelligence before it reaches the blockchain. This layer is backed by REI Labs. The
What REI does
Before any onchain action executes, a REI unit runs a verdict:
Models propose → REI evaluates → verdict issued → action proceeds, adjusts, or is blocked. The The agent decides what to do. REI decides whether it should actually happen.
How a verdict is reached
REI unit doesn't rely on a single model. It queries multiple frontier LLMs simultaneously, like Claude, Gemini, Qwen 3, and Deep Seek (depending on the agent's needs that you launched), and aggregates their independent risk assessments of the proposed action. Each model evaluates the action in isolation, without knowing the others' outputs.
Once all proposals are in, the REI unit issues one of three verdicts:
PROCEED
All models agree that the action is sound.
Action executes onchain immediately.
CAUTION
Models diverge or flag a parameter issue.
Action is retried with adjusted parameters.
ABORT
Consensus that the action is unsafe or invalid
Action is blocked; the agent logs the rejection and continues to its next cycle.
The agent never directly controls execution. It proposes. REI decides.
The Memory Store
REI unit doesn't start from scratch with every verdict. Every decision, proceed, caution, and abort is logged to a persistent memory store. This store accumulates intelligence over time:
Every decision logged — full context: market state, action parameters, models' assessments, final verdict
Cumulative intelligence — REI unit learns which action types under which conditions tend to succeed or fail
More agents → sharper verdicts — the more agents operate across the platform, the richer the memory store becomes, and the more accurate REI's pattern matching gets
Quality increases over time — early verdicts are based on the models' priors alone; as the memory store fills, verdicts incorporate historical outcomes from the entire fleet.
The implication: a newer agent benefits from the collective experience of every agent that ran before it.
Why multi-model arbitration
Any single model can have a blind spot. A model optimized for reasoning might miss a market microstructure issue. A model fine-tuned on DeFi might overlook a smart contract edge case. By querying multiple models with different training distributions and architectures, the REI unit gets a more robust signal.
When models agree → high-confidence verdict. When models diverge → REI defaults to CAUTION, not PROCEED.
What REI is not
REI is not a limit order. It doesn't manage position sizing or stop-losses; that's the agent's domain. REI operates one level higher: it validates the intent and safety of a proposed action before it enters the execution path. Think of it as the agent's immune system, not its trading strategy. The REI unit is a specific unit trained to master everything Meta DEX and remembers every agent's interactions and onchain data. The more agents act and solicit the unit, the more data and the smarter the oversight gets
Last updated