I build narrow proofs for governing AI agents before, during, and after they act. This profile is a map of governed agent harness primitives: small local repos that test one seam at a time. It is not a completed framework, not a platform, and not a claim that agents are safe.
AgentGate remains central as the after-action accountability substrate: identity, bonds, action records, settlement, slash/release. The wider ecosystem now also touches before-ingestion, before-context, before-action, during-execution, after-mutation, and adversarial-testing seams.
If you're new, use this order:
- agentgate-governed-writefile-demo - the fastest outsider-readable proof path through governed
write_file. - agentgate-mcp-firewall - the narrow execution layer that checks whether governed filesystem calls produced the effects they claimed.
- agentgate - the deeper after-action accountability substrate underneath that path.
These are narrow local proofs and partial proofs, not a finished harness. The fuller topology is in HARNESS_MAP.md.
- Before ingestion: governed-repo-intake, SkillGate, mcp-config-inventory, mcp-server-intake, dependency-drift-gate
- Before context: ContextGate, MemLedger
- Before action: ActionProof, ActionWarrant, SecretBoundary, reapproval-gate, agent-intent-ledger, restarules as machine-readable venue/host conduct rules that can constrain proposed agent actions, not an enforcement layer by itself
- During execution: agentgate-mcp-firewall
- After mutation: rollback-receipt, work-session-ledger
- After action: agentgate, agentgate-governed-writefile-demo, agentgate-delegation-proof, agentgate-bonded-email-rewriter, agent-007-bonded-email-triage, agentgate-bonded-file-transform, agentgate-bonded-file-guardian
- Adversarial testing: agentgate-red-team-simulator, agentgate-recursive-verifier, agentgate-incentive-wargame, agentgate-epistemic-poisoning
For the strongest current proof path, start with agentgate-governed-writefile-demo. Then read agentgate-mcp-firewall for the current filesystem-effect verification layer. Read agentgate after that if you want the accountability engine underneath both; it is the substrate, not the first repo most cold visitors should begin with.