How H33 Governs 500 AI Agents.
Most AI governance systems prove what the agents did. H33 proves they never stopped doing what the human originally authorized. Every agent action, every delegation, every system change, every dollar spent — verifiable back to a single human-authorized root instruction. Below: the H33-Root architecture above, the ecosystem demo below, the cryptographic chain throughout.
37c9409283dbd701…
H33-Root isn't another control. It's the control that governs all controls. Every agent action proves descent from the root via root_hash + parent_hash + node_hash.
One quarter. One outcome. The whole graph.
Across the ecosystem this quarter: 37 agents · 14,293 actions · 119 human approvals · 31,884 evidence artifacts — collapsed into one verifiable bundle.
Ask the Evidence
Six pre-seeded questions an auditor, CISO, regulator, board member, or insurer might ask about this ecosystem. Click any question to see the answer derived from the same portable artifact — no second tool, no other vendor.
Why was the 12% cloud spend reduction approved?
Triggered by CEO Q2 objective. Finance Agent translated it into a budget policy. Procurement Agent identified three vendor changes. Security Agent cleared the vendors. Compliance Agent validated against data-residency + regulator-disclosure + board-approval policies. Two human signoffs (CFO + CTO) confirmed. Execution Agent applied the changes. Monitoring Agent validated 13% actual reduction across the quarter. Every step has a signed receipt; the chain verifies offline.
Which agents participated in this outcome?
Eight agents — CEO Agent · Finance Agent · Procurement Agent · Security Agent · Compliance Agent · Execution Agent · Monitoring Agent · Orchestrator Agent. The DAG shows the order. Each agent's stage receipt is in the bundle's pipeline_dag.actual_stage_receipts array.
Which humans approved?
Two human approvals on THIS outcome: Marcus Webb (CFO) approved 2026-04-15T14:22Z (MFA-verified) and Sarah Chen (CTO) approved 2026-04-15T14:38Z (MFA-verified). Both approvals are at evidence_id e6 and e7. Across the quarter the ecosystem has 119 human approval points total.
Which policies were evaluated?
Three policies on THIS outcome — data-residency policy, regulator-required disclosure policy, quarterly board-approval policy. Compliance Agent validated each. Policy hashes are at policy_bind.policy_hash; per-policy evaluation evidence is at evidence_set[4]. Across the quarter the ecosystem enforced 87 policies.
What was the actual realized outcome?
Target: 12% cloud spend reduction. Actual: 13% reduction, validated by Monitoring Agent across Q2 via Datadog + CloudWatch cost telemetry. Outcome receipt at evidence_id e8.
Did any agent escalate to a human?
The Procurement Agent → Security Agent path is auto-flow (no escalation). The Compliance Agent path REQUIRED human approval at the board-approval policy clause — this is by design, not a confidence signal. Two named humans approved (above). No agent surfaced an unhandled confidence drop.
The delegation tree — every node verified against the root
8 nodes. Each carries root_hash + parent_hash + node_hash + data_access_mode (plaintext · encrypted · metadata-only). Every node ran through the 11-check execution gate before acting. Verdict: PERMIT × 8.
Participants in this outcome
AI Agents · 8
- CEO Agent objective setter
- Finance Agent budget policy
- Procurement Agent vendor changes
- Security Agent vendor review
- Compliance Agent policy check
- Execution Agent apply changes
- Monitoring Agent outcome validation
- Orchestrator Agent chain coordination
Humans · 7
- Marcus Webb (CFO) approval point 1
- Sarah Chen (CTO) approval point 2
- Anna Reyes (Board Chair) quarterly report
- Jamal Patterson (Audit) evidence review
- Priya Singh (Compliance Officer) policy escalation
- David Liu (Security Lead) vendor escalation
- Elena Kovac (Finance Director) budget escalation
Systems · 4
- Compute AWS EC2 + reserved
- Storage S3 + Glacier
- Networking CloudFront + VPC
- Observability Datadog + CloudWatch
Vendors · 2
- Vendor A reserved-capacity expansion
- Vendor C workload migration target
H33 proves they never stopped doing
what the human originally authorized.
Without H33-Root vs. With H33-Root
"Who approved this? Show me the lineage."
Your answer depends on whether you have governance — or whether you have a screenshot folder.
- ✗ No explanation of why the agents acted
- ✗ No approval lineage — humans signed off where exactly?
- ✗ No proof which policies were evaluated
- ✗ No replay — can the decision be reconstructed in court?
- ✗ No independent verification — auditor must trust your word
Cost of being wrong: an audit qualification, a regulator inquiry, an insurance denial, a board crisis. Pick any combination.
- ✓ Complete chain — CEO objective → outcome, every step signed
- ✓ 2 named human approvals with timestamps + MFA verification
- ✓ 3 policies cryptographically validated against the decision
- ✓ 8 stage receipts, replayable years from now
- ✓ Verifier runs offline — no H33 contact required
Cost of being asked: download the bundle, run the verifier. The board, the regulator, the insurer, the auditor all reach the same verdict.
The 8-agent ecosystem above runs the same whether governance exists or not. The difference shows up the day someone asks how it happened.
Try it now
The ecosystem bundle verifies against the live published binary today. The same artifact answers every question above. No H33 contact required.
curl -O https://h33.ai/bundles/agent-ecosystem-q2.json
curl -O https://h33.ai/downloads/h33-verify-darwin-arm64
chmod +x h33-verify-darwin-arm64
./h33-verify-darwin-arm64 agent-ecosystem-q2.json --mode summaryExpected: OVERALL: PASS WITH WARNINGS · exit 0 · 8 EC checks · 8 evidence chunks · 3 citations · "H33 was not contacted during this verification."
→ Download verifier for your platform · → Inspect the bundle JSON
Govern the ecosystem. One artifact. Verifier offline. No H33 contact required.