Your agent
Agent safety
An agent can't widen access, invent evidence, ratify as a person or turn a proposed action into proof of execution.
Your agent works within the access and authority of its connection. A prompt doesn’t widen either one.
What an agent can never do
Section titled “What an agent can never do”- Read across workspace, boundary, compartment or sensitivity restrictions.
- Reveal the existence of records you’re not cleared to know about.
- Invent a source, quote, citation, signature or receipt.
- Turn its own summary into your verbatim words.
- Ratify a decision as a person or treat silence as approval.
- Bypass a failed evidence check with another writing route.
- Execute an external action merely because it proposed one.
- Erase an earlier entry to hide a correction.
An agent credential doesn’t prove a person is present. A browser sign-in identifies the person. A consequential approval still needs the relevant authorization.
Check the result
Section titled “Check the result”Ask for the stored record and receipt after a write. Read the action status after an approved action. If the agent can’t retrieve the result, it should say what’s unverified.
When a tool is unavailable, review the connection and grant. Don’t give the agent another person’s credential or paste a secret into a conversation to work around the restriction.
Planned routine controls
Section titled “Planned routine controls”Delegated routine grants are coming soon. They’re limited to the agreed work, revocable and never wider than your access. They never permit ratification or workspace administration. See Sweeps and routines.
Syderial is the system of record for intelligence.