1. Govern the software factory, not just its output
The Governed Software Factory Loop connects agent-assisted delivery to evidence and review. The useful sequence is plan, execute, test, retain evidence, evaluate risk, approve and promote. Passing tests is not sufficient when ownership, traceability or rollback is missing.
Harness delta: L2 promotion gates, L3 orchestration, L7 evaluation and SDLC traceability.
2. Test security for fail-open behavior
DefenseClaw System One was incorporated as benchmark evidence emphasizing fail-open cases. Security tests should examine missing policy decisions, ambiguous tool results and degraded guardrails, not just successful responses. A benchmark is a test instrument, not proof of deployment safety.
Harness delta: L6 deny-by-default enforcement and L7 negative and failure-mode evaluation.
3. Treat safety controls as independent barriers
The IOGP Process Safety Fundamentals integration supplies a useful engineering analogy: effective barriers matter more than confidence in an operator. For agents, separate model judgment from identity, authorization, execution limits, monitoring and recovery. Industrial process-safety guidance is not an AI certification.
Harness delta: L6 preventive controls, L7 barrier-effectiveness evidence and L8 isolation and recovery.
4. Consolidate discoveries without losing provenance
A catalog consolidation brought REA, MASS and openTPU work together while preserving individual source records and index order. Discovery should remain broader than promotion. Every candidate needs source, version, license, limitations and an explicit integration decision. Catalog inclusion does not mean deployment.
Harness delta: L4 provenance, L5 capability catalog and L2 controlled promotion.
5. Make repository hygiene enforceable
The English-only language gate and related audits turn an editorial convention into testable policy. Machine-readable findings and CI checks improve metadata consistency. Language conformity complements evidence review; it does not prove technical correctness.
Harness delta: catalog CI, reviewable source metadata and integrity checks.
What changes next
Bind every promoted change to a source version, an L0–L8 control mapping, adverse-condition tests, retained evaluation artifacts, a named approver and repository-wide consistency checks. An agent may propose an improvement; it must not approve its own authority or silently change the evidence used to qualify it.
Bottom line: a trustworthy agentic system is not one that never fails. Its authority is bounded, failures are observable, evidence survives the workflow, and promotion is deliberate.
Evidence scope: selected Harness-collection commits and merged pull requests reviewed on 18 October 2026. External findings retain their own provenance and qualification status.
Privacy Notice: With your consent, this website uses Google Analytics to understand aggregate visits and content usage. Analytics is optional and is not used for advertising or cross-site profiling. Some content and images may be generated or modified with the assistance of AI; every AI-assisted item is reviewed and manually edited by a human before publication. Full Privacy Policy ·