Enforced, checked, or a person
How Assent will say what it can and can’t promise: three kinds of assurance, each with its limit, in every report.
Design noteThis note describes how Assent is designed to work. Intent, not results.
- 01What the runtime can enforceSome constraints can be enforced while an agent works.
- 02What a check can establishSome can only be checked afterwards.
- 03What needs a personSome need a person’s judgment.
Three kinds of assurance
Some constraints can be enforced while an agent works. Some can only be checked afterwards. Some need a person’s judgment. Treating them as one kind of assurance is how false assurance happens, so the design keeps them apart, and claims grow only after new coverage is verified.
Source: Decision and assumption register
What the runtime can enforce
Because Assent would operate the runner, restrictions can come from the runtime itself: operating-system or container limits on the paths an agent may write, provider keys held outside the sandbox, outbound traffic on an allowlist, no production credentials, and separate identities, so the agent can’t alter policy or its own attestations.
Each has a limit. Where a runtime can’t restrict a path, the final diff check would catch the write after the fact; it would not prevent it. A missing permission is not proof of isolation, so the environment itself would be tested. An agent that reads untrusted content can still be misled, and exfiltration through allowed channels can’t be fully excluded. And instructions kept in the repository can’t protect themselves.
Source: Authority and enforcement
What a check can establish
After a run, the candidate API definition would be compared with the approved contract, and compatibility rules and consumer tests would run. The results would be bound to the exact candidate and base commits and published by a trusted publisher, separate from the runner.
The limits would stay attached: a structural comparison doesn’t detect every runtime behavior change, and passing results cover defined rules and exercised behaviors only.
Source: Authority and enforcement
What needs a person
Transaction and security semantics would go to explicit human review. New consequential unknowns would be raised by the agent’s report and by repeated-failure and structural signals, then reviewed independently, because detectors can miss concerns and self-report isn’t a complete boundary.
Source: Authority and enforcement
Saying it, every time
Every report would show failed, unavailable, passed and unverified constraints separately, and completion would always list what remains unverified. An implementation can pass compatibility checks and still need review for transaction, security or behavioral semantics.
Source: Product workflow
Small samples, stated as small
Unavailable checks would never be counted as passes. And zero misses in a small sample is not proof of zero risk; results will say so.
Source: Metrics and economic evidence
Design partners
Bring one recent cross-team change.
We reconstruct it with you before showing a solution: the decisions, their owners and the effort, agent operation included.