# Governance Governance is not a policy PDF beside the system. It is the manifest, the gate, the held state, the credential boundary, and the event record inside it. Source: https://aioverload.io/governance/ {{< section id="manifests" eyebrow="DECLARED STATE" title="Access changes as a reviewed diff." >}} Agents, workflows, providers, and grants live in a registry. Git and the forge provide the review surface operators already use; a merged change is what changes desired access. This makes a grant inspectable before a run inherits it. ```yaml workflow: dev-review-loop-reflex goal: reviewed, verified change acceptance: - tests: pass - reviewer: PASS tools: declared-only ``` The snippet is illustrative, not a complete engine manifest. Its purpose is to show the governance unit: purpose and permission travel with the work. {{< /section >}} {{< section id="gate" eyebrow="ADMISSION" title="A risky action becomes a state, not a surprise." class="section--tint" >}} A governed step can dispatch, hold, or block. When the reflex score crosses its deterministic threshold, the run remains in `step.held` with a reason attached. An unavailable gate fails closed for gated work.
GODispatch

The step is within policy.

HOLDWait for a human

Risk or uncertainty needs judgment.

BLOCKFail loudly

The requested action is refused.

The model scores; the deterministic rails decide. This separation keeps a generative system from talking its way around a threshold. {{< /section >}} {{< section id="approvals" eyebrow="HUMAN JUDGMENT" title="Approval is structured input, with a name and scope." >}} An operator can approve with constraints or reject the held step. The decision and the operator’s words land on the same event log as the work they authorize. Approval does not erase the risk event that caused the pause; it completes the record. ```bash aioctl approve review_2 \ --input "review ok — do not push" ``` This is narrower than a generic “continue” button: the input can constrain the next dispatch and remains visible during replay. {{< /section >}} {{< section id="tools" eyebrow="LEAST PRIVILEGE" title="The tool plane starts at no." class="section--dark" >}} Gateway-routed MCP calls are limited to registry-allowlisted servers and the run’s declared grants. Per-run credentials reduce the lifetime and scope of supported provider access. An undeclared target is denied at the gateway rather than merely omitted from a prompt. **The boundary is not universal yet.** Harness-native tools that bypass the gateway are not covered by this gate, and subscription sessions can inherit credentials. The project names those gaps in its [security model](/security/#gaps) instead of implying complete isolation. {{< /section >}} {{< section id="evidence" eyebrow="ACCOUNTABILITY" title="Every governance decision leaves the same kind of evidence." >}} | Decision | Recorded evidence | |---|---| | Definition changed | Reviewed registry or workflow change | | Step held | Risk score, threshold, reason, and held event | | Operator approved | Actor, input, scope, and approval event | | Tool refused | Gateway decision and requested target | | Run completed | Ordered events, outputs, and definition hash | Replay turns that evidence into a verifiable reconstruction. Governance is therefore not only what the system refuses; it is what the system can prove afterward. {{< /section >}} {{< next "/security/" "Read the threat model" "See the containment layers, the gaps the gate does not cover, and the disclosure path." >}}