An audit record for every step an AI agent takes.
Each agent gets its own identity and only the access you grant. You can check any step and stop any agent.
Today this is a sandbox, and it only controls agents that HIL runs itself. Controls for agents your firm already runs, starting with Microsoft Copilot Studio, are designed but not built.
- agent
- did:web:hil.systems:aid:example
- step
- 3 read_file claims/2026-q3-summary.pdf
- hash
- 03acc6b91d9d… follows ef838629e8f1…
- sig
- ed25519 MNiV6ryGSgLGIlOH…
- check
- signature verified
Try the two things auditors ask about
Both of these run in your browser. The first uses real SHA-256 hashes, the same way HIL's log does.
Can anyone quietly change the record?
audit log, 5 entries
- 1
Agent 7f3c signed in with its own key
hashing…
- 2
read_file claims/2026-q3-summary.pdf
hashing…
- 3
write_file briefing.md (4,812 bytes)
hashing…
- 4
draft_email to j.nguyen for review
hashing…
- 5
Task accepted by reviewer
hashing…
Every entry carries the hash of the one before it. Try editing one.
Can a person stop an agent mid-task?
agent 7f3c running
- list_files in the shared workspace
- read_file claims/2026-q3-summary.pdf
- read_file policy/complaints-handling.md
- write_file briefing.md
- draft_email to j.nguyen for review
- Submit output for review
Press and hold for a moment. It stops at its next step.
HIL in 25 seconds
APRA's letter, a signed step, an edit the chain catches, and an agent stopped mid-task. Made from HIL's own demos, which use real hashes and signatures.
Who uses it at your firm
Each of these people can answer a question about an agent by looking at the record.
Every agent is listed with the person responsible for it, what it is for and what it may touch. You can stop an agent or end its engagement, and its work halts at the next step. Only the issuer's officers can revoke an identity.
| Agent | Responsible | Access | |
|---|---|---|---|
| Claims summariser | Aisha Karim | read, write | Active |
| Complaints drafter | Tom Nguyen | read, draft | Active |
| Policy checker | Ben Walsh | read | Active |
| Rates assistant | Mei Chen | revoked | Revoked |
APRA wrote to every bank, insurer and super trustee about AI on 30 April 2026.
Identity and access management “have not yet adjusted to nonhuman actors such as AI agents”.
From 10 December 2026 your privacy policy must also say when a computer program makes, or substantially helps make, a decision that could significantly affect someone. For agents HIL runs, the record shows which agent did what, for whom and when, so you can answer both from one place.