Agent permissions
The agent inventory: identity, tool permissions, approved repositories and environments, risk class, kill switch and an honest enforcement level.
Each agent in the inventory records its provider, identity, tool permissions, approved repositories and environments, and risk class. A kill switch stops the gateway from allowing anything for that agent.
Enforcement levels
Monitored: TrustFabric only sees what is reported. It cannot block the agent.
Voluntary: the agent calls the gateway through a hook it could disable, such as the Claude Code PreToolUse hook script.
Enforced: the execution path is mediated, such as a CI gate step that blocks the pipeline, or a Claude Agent SDK agent whose tool calls pass through can_use_tool in your own runner.
GitHub Copilot and OpenAI Codex are monitored or voluntary only. No mediated path is claimed for them.
What agents can never do
Agents cannot approve their own requests, change policy or mint approvals. Those are human-only permissions, and workload tokens never hold them.