Governance

An agent company has to answer the questions a regulator, an auditor or a board asks about any workforce: who works here, on whose authority, doing what, watched how.

The registry. Every agent has a signed card (.well-known/agents/) — identity, purpose, model, tools, memory stores, egress, cell, autonomy, the human role it answers to. Cards are rendered from the manifests and checked against them in CI, so a card cannot drift from the agent it describes.

Autonomy and oversight. Each agent declares an autonomy level; human gates on pipeline stages are where people decide; the Desk is where those decisions are made and recorded. Kill switches stop an agent, a stage, a pipeline, a cell or the workloads.

The audit log. Every console action, with who did it (a user, an app key acting as its user, or the Human), what it changed and what it was before. Revert where reversing is possible.

Observability. Per-call traces of every run for thirty days; per-workload draw on the Seats and models facet; the evals matrix says what each model has been judged at on each workload.

Evidence. Governance evidence packs, quarterly attestations and the mapping to NIST AI RMF and EU AI Act tiers arrive with the Governance board (programme PR 11). Pro.

The Governance board

Open the Governance lens on any level of the map (key 7) and press Governance board. It is the registry as a board, and nothing on it is typed in twice: every record is derived from the manifest, the signed card, the workloads, the run ledger, the vault grants, the egress allowlist and the halt file.

• One record per agent — identity (the card's signature state), the human role it answers to and the Steward and Asset Owner users bound to it, purpose and use case, model, tools and MCP servers, memory stores and vault mounts, egress, cell, lifecycle, risk tier and autonomy, last attestation, and its allowance draw over thirty days with an anomaly flag.
• Risk tiers come from rules/governance.yaml: minimal, limited, high, unacceptable, each mapped to the EU AI Act's tier and the NIST AI RMF functions the controls have to satisfy. A manifest may set risk_tier and use_case; otherwise the tier follows the agent's capability words.
• Policy as data — an agent's autonomy, tier and use case can be overridden per cell, always shown as a diff against the cell default, with the role that approves each action class at that autonomy (rules/approvals.yaml).
• Quarterly attestation — per cell, the Steward (or the Human) confirms every agent is still meant to exist; a due attestation is a card on the Desk, and an overdue one turns the cell's boxes red under the lens.
• Kill switches at five scopes — agent, stage, pipeline, cell, everything — all written to the one halt file every run reads before it starts.
• Shadow agents — anything that ran or is loaded without a manifest.
• The evidence pack — for any period, as JSON or Markdown: the registry, attestations, policy and kill-switch changes, decisions, runs per agent and each control with where the code enforces it. Also over MCP (list_agents, agent_policy, evidence) and at /api/governance/evidence?from&to&cell.