Skip to main content

Governance and audit for AI agents, MCP servers, and workflows

AgentPaaS connects workload identity, policy decisions, run records, and audit evidence. This page explains how governance works across agents, MCP servers, tools, and workflows, then links to the command procedures and verification details.

What AgentPaaS governs​

During a governed run, AgentPaaS records the workload, deployment, run, policy, and request path. The platform can record:

  • Allowed and denied egress decisions.
  • Declared destinations and credential bindings.
  • Model and tool calls.
  • Parent and child request lineage.
  • Workflow stage progress and member runs.
  • Signed package and deployment identity.

These records help you investigate a run and review whether it followed the declared policy. They do not prove that the workload code is safe or that a prompt injection cannot occur.

How audit evidence is created​

Every invocation creates a run. AgentPaaS connects the run to its deployment and records execution and policy events. The audit path uses hash-chained JSONL records. Signed export bundles can add checkpoint signatures.

Use Runs and Audit Logs to inspect a run, read logs and events, and export records. Use Audit export and verification to verify an export on a second machine.

Governance for MCP servers and tools​

MCP servers and tools use the same deployment controls as other components. Their declared destinations, credential bindings, invocation records, and policy decisions remain part of the governed request path.

See MCP servers, Tools, and Credentials and secrets.

Governance for workflows​

A workflow is a signed recipe that connects deployed components. The workflow record identifies its stages and member deployments. Each invocation creates a workflow run with stage progress and member runs.

See Workflow kinds and edges, Workflows, and Known limitations.

Review workflow​

  1. Inspect the deployment, declared policy, destinations, and credential bindings.
  2. Run the agent, MCP server, tool, or workflow.
  3. Review the run status, result, logs, and events.
  4. Export the audit records for the run.
  5. Verify the export on a second machine.
  6. Record any policy decision or unexpected request that needs follow-up.

The security review page links to the threat model, enforcement details, data handling, compliance status, and current limits.

What verification detects​

Audit verification detects modified records, reordered records, and inserted records in the middle of a hash chain. In the current release, deleting the final records from an exported file can leave a valid prefix chain. External checkpoint anchoring is required to detect that tail deletion. See Known limitations.