Logs every agent action, keeps it forensically separable from human actions, and human-gates mutations.
v3 · Memoise + batch · Jun 2026 – Memoise identity lookups; batch low-risk approvals (deltas vs v2)
Dashed tiles are the markdown / JSON context we author around the harness.
Integration logic – MD files
mutation-audit.mdSeparate agent vs human mutations and flag any missing a gate
policy-gate.mdEvaluate an action against policy and route the approval
identity-resolution.mdResolve the acting principal for every logged action
siem-anomaly-scan.mdScan security events for anomalous agent behaviour
autonomy-promotion.mdCheck the criteria to promote an agent from L2 to L3
The model is rented and swappable. This is the part we own and port: the MCP tools, the markdown / JSON context wrapped around them, and the systems they connect.
audit_log_writeAppend every agent & human action
servicenow_approvalCreate & route human-approval requests
splunk_queryQuery Splunk security events
opa_evalEvaluate an action against OPA policy
okta_lookupResolve the acting principal via Okta
governance/audit-rules.mdWhat is logged & how actors are separated
governance/gates.jsonWhich actions require a human gate
governance/policies.mdCriteria to promote autonomy L2 → L3
Swap freely – the harness above doesn't change.