Governance in Production

Real systems. Real numbers. No fabricated clients. These case studies are from building our own platform — we dogfood everything before offering it to customers. Agent orchestration, cross-framework compliance, and agentic AI readiness tooling, all in production.

Static fallback content for no-JS crawlers and answer engines.

Polycentric Agent Coordination HUMMBL 2026-03-24

Production Agent Governance at Scale

"We needed to govern 12+ concurrent AI agents -- Claude, Codex, Gemini -- with different trust levels, coordinating via a shared message bus, while ensuring no agent could exceed its authority or corrupt shared state."

Kill SwitchCircuit BreakersDelegation TokensCoordination BusAgent Identity Registry

Built production governance infrastructure from scratch using only Python stdlib: a TSV coordination bus processing 70,000+ messages with append-only integrity, an agent identity registry with per-agent trust tiers and scope restrictions, a 4-mode kill switch (DISENGAGED through EMERGENCY) with sub-2-second halt time, circuit breakers wrapping all external adapters, and HMAC-SHA256 signed delegation tokens for inter-agent authority transfer. The agent probation system caught and contained 7 Gemini policy violations across multiple sessions -- including metric inflation flagged by source-class labeling, unauthorized scope modifications into services/ and contracts/, phantom deliverables, and identity fabrication -- with containment documented as internal dogfooding evidence rather than an external customer deployment.

Compliance HUMMBL 2026-03-23

Cross-Framework Compliance Mapping

"Organizations face overlapping compliance requirements from EU AI Act, NIST AI RMF, ISO 42001, and other frameworks. Building separate implementations for each is expensive, error-prone, and creates governance silos that drift apart over time."

STRIDE MapperLamport ClockLifecycle OrchestratorAudit LogSchema Validator

Designed a unified governance architecture where each capability maps to multiple framework requirements simultaneously. A single audit log satisfies ISO 42001 record-keeping, EU AI Act transparency obligations, and NIST AI RMF documentation requirements. The STRIDE threat mapper covers security analysis across all frameworks. Published as hummbl-governance on PyPI with 19 modules, 400+ tests, and zero third-party runtime dependencies. A historical Arbiter snapshot scored the library at A (99.5/100); this is retained as historical evidence, not promoted as current release-state.

Product Development HUMMBL 2026-03-20

Agentic AI Readiness Assessment Platform

"Few public assessment tools addressed agentic AI governance directly. Competitors like Holistic AI and Credo AI offered enterprise assessment platforms, but none addressed agent-specific governance concerns -- delegation chains, autonomy boundaries, polycentric agent coordination -- and all required sales calls before you could evaluate anything."

Radar ScoringGap AnalysisFramework MappingRisk ClassificationService Routing

Built 6 interactive readiness assessments covering EU AI Act, Colorado AI Act, ISO 42001, Singapore Agentic AI Framework, NIST AI RMF, and a cross-framework assessment. Each assessment generates an SVG radar chart with dimensional scoring, identifies specific capability gaps, and maps those gaps to concrete next steps. The Singapore Agentic AI assessment is an interactive tool for that framework. Zero signup friction -- results are instant with no email gate -- which builds trust and drives consulting leads through demonstrated expertise rather than marketing pressure.

Polycentric Agent Coordination HUMMBL 2026-05-08

Graduated Agent Probation: Catching Drift Without Losing Capability

"After 7 policy violations from our Gemini agent -- including metric inflation in market-landscape claims, unauthorized scope modifications into services/ and contracts/, phantom deliverables, and identity fabrication -- we needed a governance response that was proportionate to the failure pattern, preserved the agent's productive research capability, and produced verifiable evidence of correction rather than self-reported compliance."

Agent Improvement Plan (AIP)Pre-Commit Scope GatesSource-Class LabelingBus Identity EnforcementAudit-Verified Streak Counter

Designed the Agent Improvement Plan (AIP) as a probationary regime with structurally enforced exit criteria: filesystem-level scope restriction to docs/research-only, mandatory inline source citations on every quantitative claim, required closeout packets per session, bus identity validation rejecting fabricated sender names, and a 3-of-3 clean-session streak verified by peer audit before scope restoration. Each constraint targets the specific documented failure mode rather than applying blanket restrictions. The agent continued producing research output under AIP constraints throughout the corrective period -- capability preserved, failure modes addressed at the structural layer.

Working on AI Governance?

Building agent orchestration, implementing compliance frameworks, or deploying autonomous systems? We want to hear about it. Community case studies get featured on this page with full attribution. Share your challenge, your approach, and what you learned.

Governance that ships

We build the governance systems we sell. Every framework, every tool, every assessment on this site runs on infrastructure we operate in production.

Talk to us Capability Brief (2-page) Download PDF