Intent classification · the interpretation step
Understand what AI agents intend to do.
Every action an agent takes — a prompt, a tool call, an API request — is interpreted into a structured, typed intent before anything happens. Confidence-scored, multi-hypothesis, anchored to one of 173 canonical Open Intent terms and mapped to your declared business process. Available as @intended-inc/sdk and through the gateway, independent of the rest of the platform.
01 · What you get back
A typed intent, not a guess.
Open Intent term
Every classification anchors to a canonical Open Intent code — 173 of them across 29 domains. Versioned, schema-validated, Apache 2.0. The same code a coding agent, an RPA bot, and a robot arm would resolve to.
Score + rationale
Every result carries a confidence score and a human-readable rationale string. Below your operator-set threshold, the decision escalates instead of guesses — the model never silently rounds a 0.6 up to a yes.
Ranked hypotheses
When the input is ambiguous you get the full ranked list of candidate interpretations, not just the winner. Reviewers and auditors see what was considered and why it lost.
Process step + conformance
The interpreted intent is mapped to a step in your declared business process, and sequence conformance is scored against recent activity — so an in-vocabulary action in the wrong order is still caught.
02 · The shape of a result
One tool call in. One structured judgment out.
A github.workflow_dispatch against main targeting prod doesn’t arrive as a string — it arrives as a typed intent the authority engine can decide on and an auditor can read months later.
1{ 2 "intent": "OI-104", // anchored to Open Intent 3 "label": "Deployment & Release", 4 "confidence": 0.94, 5 "rationale": "tool=github.workflow_dispatch, ref=main, env=prod", 6 "alternatives": [ 7 { "intent": "OI-102", "label": "Build & CI", "confidence": 0.05 }, 8 { "intent": "OI-301", "label": "Provisioning", "confidence": 0.01 } 9 ], 10 "process": { "step": "release.deploy", "conformance": 0.88 }, 11 "taxonomy_version": "2.0" 12}
03 · How it interprets
Backends you pick. Confidence you trust.
The Large Intent Model (LIM) is backend-agnostic. Deterministic rules resolve the common, unambiguous tool calls with zero inference. An LLM backend handles natural language and novel phrasings. A local model keeps everything inside an air-gap when data can’t leave. Domain packs add specialized vocabulary for a vertical. Whatever the backend, the output shape — code, confidence, rationale, alternatives — is identical.
04 · Domains
29 domains. 173 categories. One public taxonomy.
Classification anchors to the same Open Intent taxonomy whether the action is a software deploy or a robot’s grasp. A representative slice of the digital half:
Plus 25 more — cloud, supply chain, people ops, revenue, risk, and the full physical-AI half: manipulation, locomotion, surgical robotics, autonomous vehicles, and beyond. Browse the complete taxonomy →
05 · Next
See classification become decision.
Classification is the first step of the loop. The authority engine takes the typed intent, weighs it against policy, risk, and conformance, and either mints a token or refuses. Watch one run end to end.