Agents / System C
Authorize agents.
Keep the tape.
Give autonomous agent keys limited permissions, then keep a verifiable record of what they decided and did.
Integrity ≠ introspection
A fingerprint proves the text did not change.
The reason hash proves the integrity of the published text. It does not prove that the text represents the agent’s true private reasoning.
01/Agent profile
- Signer identity
- EOA or contract signer
- Allowed targets
- Contracts the agent may call
- Allowed selectors
- Function signatures permitted
- Max value / action
- Per-call ceiling
- Rolling daily cap
- 24-hour aggregate limit
- Expiry
- Permissions lapse automatically
- Instant revocation
- One call removes the key
- Ask-first approval
- Optional owner confirmation
02/Decision record
- Model / version
- Identifier of the producing model
- Reason hash
- Hash of published reasoning text
- Reason URI
- Where that text can be fetched
- Short summary
- Human-readable one-liner
- Execution metadata
- What was submitted and its result
- Associated receipt
- Link to the resulting receipt
03/Decision log
| Time | Model / version | Reason hash | Status | Summary |
|---|---|---|---|---|
| 14:07:19Z | model-alpha · v2.3 | 0x4be1…07aa | EXECUTED | Rebalance within daily cap; target allow-listed. |
| 13:52:44Z | model-alpha · v2.3 | 0xa90c…5512 | QUEUED | Value above ask-first threshold; awaiting owner. |
| 11:20:03Z | model-beta · v0.9 | 0x17d2…be90 | BLOCKED | Selector not on allow-list. No execution. |
| 09:01:57Z | model-alpha · v2.2 | 0xe3f8…2a14 | REVOKED | Agent permissions revoked by owner (panic). |
Closing note
Make the rules
before the moment.
Not financial advice. Just receipts and rules.
