Use case · DevOps · OI-100
Coding agents, governed at the gate.
An autonomous SDLC agent can open a thousand pull requests, trigger a hundred deploys, and rotate every key in a vault before a human reads the first alert. Intended verifies what each action intends to do — against your declared release process — and gates it with a single-use Authority Token the CI runner verifies itself. No token, no deploy. The whole 7-category Software Development domain maps to one canonical vocabulary.
01 · Why the deployment boundary
Branch protection assumes a human pressed merge.
Every SDLC control was built around a person in the loop. Required reviewers assume someone read the diff. The change ticket assumes someone filled it out. A coding agent with a valid GITHUB_TOKEN satisfies all of those without a single human judgment — the credential says can, never should. Intended re-introduces the missing judgment as a runtime gate: it interprets the action, anchors it to a canonical intent, scores it against your declared release process, and decides before the runner moves.
02 · The OI-100 surface
Every SDLC action resolves to one of 7 canonical intents.
The Software Development domain (OI-100) is the shared language a deploy speaks whether it came from a coding agent, a CI bot, or a human running a release script. A policy written against a code keeps working across taxonomy versions — codes are stable family identifiers.
Code Management
Build & CI
Testing & QA
Deployment & Release
Change Control
Architecture & Design
Package & Dependency
03 · What gets gated — and how the risk scores
Production deploys
Token-bound workflow_dispatch triggers. Verified inside the GitHub Actions runner; the deploy job has needs: verify and never starts without an ALLOW. Production adds an environment risk modifier.
Merges to protected branches
PR-merge authority is conditional on conformance score, composite risk threshold, and the configured approver pool. A force-push that rewrites protected history scores as high-risk on inherent severity alone.
Schema migrations
Irreversible or data-loss-class migrations carry elevated inherent risk; the production environment modifier compounds it. The decision rationale — every contributing factor — is sealed into the audit chain.
Rotations & secrets
Vault-agent and similar rotators acquire a single-use token per rotation. Out-of-window or bulk rotation raises the velocity signal before the action executes.
04 · The connector decides, not the agent
Verification lives in the runner.
The GitHub Action connector verifies the Authority Token inside CI — signature, TTL, intent hash, revocation — and fails closed if anything is off. The guarantee survives even if Intended is offline, because the connector checks the token against your tenant’s kid-pinned public key, not a live API call. Per-tenant JWKS verification is live.
1# Illustrative — packaged Action is on the roadmap. 2# Today: run a step that calls the Intended API / SDK. 3- on: workflow_dispatch # agent triggers the production deploy 4 jobs: 5 verify: # → POST /v1/agent-intent → decision + Authority Token 6 action: "Deploy to production" # OI-104 · Deployment & Release 7 deploy: 8 needs: verify # no Authority Token → job never runs 9 if: steps.verify.outputs.decision == 'ALLOW'