intent

Unaudited. Advertised on trust: no gate has verified that the runtime does what this page says. Treat it as documentation, not as a guarantee.

Since v0.1.0 · Top-level declaration

Grammar

intent <Name> {
    given: <Identifier>         # optional — typed input bound to this intent
    ask: "<prompt>"             # optional — natural-language target
    output: <TypeExpr>          # optional — declared output shape
    confidence_floor: <0.0..1.0># optional — minimum confidence to accept
}

intent declares what an agent is meant to achieve, independent of how the achievement is decomposed into steps. Where flow is the procedural composition (the recipe), intent is the declarative target (the dish). Both can coexist — adopters write a free-form intent alongside a flow so the audit trail records the operator's stated objective beside the mechanical decomposition.

intent is the cleanest place to express "I want X" without yet committing to the steps. The runtime stores the intent verbatim in the audit chain; reviewers can later audit whether the flow's actual output addressed the declared intent.

Surface

intent is a top-level declaration. It is not nested inside a flow or step.

intent SummarizeContract {
    given: doc
    ask: "Produce a one-page executive summary of the contract."
    output: ContractSummary
    confidence_floor: 0.8
}

Fields

given: (optional)

A single identifier naming the typed input the intent operates over. Typically a flow parameter name. If the flow that addresses this intent declares (doc: Contract), the intent's given: doc records the binding.

ask: (optional)

A string literal containing the natural-language objective. This is what the auditor reads to know "what was the operator trying to do?". Unlike a step's ask: (which the model consumes as a prompt), the intent's ask: is declarative — it states the target outcome, not the instruction to produce it.

output: (optional)

The declared output type the intent targets. Accepts the full type-expression shape (bare types, generics, optionals, nested generics like FlowEnvelope<List<Risk>>).

confidence_floor: (optional)

A numeric literal in [0.0, 1.0]. The minimum confidence threshold the intent demands. If the flow's actual output drops below this floor, the runtime emits an intent_confidence_violation audit row regardless of whether the binding flow's anchors triggered.

Runtime behaviour

intent does not by itself execute anything. It is a declarative artifact recorded in the audit chain. The lowering creates an IntentNode IR entry; the runtime cross-references it with the flow's execution trace and emits a session:intent_review row at completion that pairs (intent.ask, flow.output, intent.confidence_floor, flow.measured_confidence).

This pairing is what lets a downstream auditor answer "did the agent address the stated intent?" without having to reverse- engineer it from the step bodies.

What this primitive is NOT

• Not a flow. A flow is the procedural composition that addresses an intent; the intent declares the target.
• Not a step. A step is an operation; an intent is a stated outcome. The two operate at different layers.
• Not an anchor. Anchors are typed predicates evaluated at emission time. An intent is a declarative artifact reviewed after the fact.
• Not required. A flow can run without a declared intent. Adopters declare one when the audit trail's "what was the goal?" answer matters.

See also

• axon://primitives/flow — the procedural counterpart that addresses an intent.
• axon://primitives/anchor — runtime-enforced predicates; complements the declarative intent.
• axon://primitives/step — the unit of cognition inside a flow.