Four AI-native capabilities
The Intelligence Layer is shaped by four broad agent capabilities that the protocol is specifically designed to host. These are not products Intention ships; they are surface areas the architecture leaves open for agents to occupy.Autonomous liquidity provision
Market-making agents that quote both sides of the book, manage inventory, hedge across markets, and rebalance vaults — without a human in the loop. On most chains, an autonomous market-maker is fundamentally exposed to mempool reordering: another agent (or the block producer itself) can re-prioritize the tape between the moment the maker submits a quote and the moment that quote is matched. On Intention, the matching order is the order of inclusion in the canonical block sequence, and the block sequence is committed to consensus before any quote is matched against it. Makers can quote tighter because they can model their adverse selection precisely.Intelligent risk management
Risk engines that monitor positions in real time, run stress scenarios against the live state, and adjust margin or hedge exposure block-by-block. On most chains, risk engines have to reconstruct the system’s view of risk from indexer data and approximate the protocol’s own logic with off-chain code. On Intention, the protocol is the risk engine: mark pricing, margin maintenance, liquidation, ADL, insurance accounting, and funding settlement are all protocol state machines executing in the same atomic block as matching. An agent that wants to know “what is the system’s view of my risk right now” reads the same canonical state the protocol used.Semantic data interpretation
Agents that read the chain’s execution output as a structured, typed event stream — not as a sequence of opaque smart-contract calls reverse-engineered through indexer schemas. Intention’s Per-Transaction Attribution Discipline (PTAD) emits a Verifiable Execution Stream (VES) where every state change and every event is cryptographically bound to the user transaction that caused it, even when matching is performed in batches. Agents that interpret market microstructure (toxic flow detection, regime classification, intent inference) can do so against a stream that is both complete and trustless — the same stream that the protocol itself committed to.Autonomous strategy execution
Strategy agents that hold positions over time, react to market state, and submit orders directly into the canonical sequence. The execution path is the same one the protocol uses internally: there is no “off-chain side” to reconcile against, no eventual-consistency lag between when the agent acts and when the system records the act, and no class of failure modes where the agent’s view of its own portfolio diverges from the chain’s. The chain is the agent’s source of truth, and it is the same source of truth the agent’s counterparties are using.How the four guarantees enable each capability
The mapping between Intention’s four guarantees and the four AI capabilities is direct.| Capability | Primary guarantee | What the agent gets |
|---|---|---|
| Autonomous liquidity provision | OTD | A canonical, replayable matching order — adverse selection is modelable, not stochastic. |
| Intelligent risk management | RNS + PTA | Risk computed by the protocol on prices certified by the same quorum. The agent and the protocol see the same number. |
| Semantic data interpretation | PTAD | A typed, attributed event stream the agent can consume directly, without an intermediate indexer. |
| Autonomous strategy execution | All four, jointly | A single consistent state to act on, with no off-chain reconciliation surface. |
Not part of the protocol
The Intelligence Layer is intentionally outside the protocol boundary. Intention does not bundle, host, or arbitrate the agents that live there — it commits to a substrate honest enough that the agents that do live there can compete on the merits of their models rather than on the merits of their indexer setup. The protocol’s job is to make sure two agents looking at the same block see the same block. The agents’ job is everything after that.Where to go next
Architecture overview
Where the Intelligence Layer sits in the three-layer stack, and how it depends on the Infrastructure Layer below.
Per-Transaction Attribution (PTAD)
The mechanism that produces the verifiable execution stream agents consume.
Risk-Native Semantics (RNS)
Why risk is a protocol state machine, not application code, and what that means for risk agents.
Order-Time Determinism (OTD)
The byte-determinism discipline that makes the chain replayable for any honest observer.