# Try the agent access review

Rethinking Agents · v0.2 · September 17, 2026

Resource: https://rethinkingagents.com/resources/agent-access-review
Blank record: https://rethinkingagents.com/resources/agent-access-record.md
Worked example: https://rethinkingagents.com/resources/agent-access-worked-example.md

## A 20-minute review, without private systems

This is a usability exercise, not a security test or an interview already conducted. All organizations, repositories, configuration observations and workflow details below are fictional. No commands, credentials or actual agent are needed. The time limit is a proposed testing target.

1. Before using the resource, note what you would normally check before enabling this workflow.
2. Read the fictional packet below, then use the resource and blank record. Decide: enable with constraints, defer for specific evidence, or keep manual. Record evidence and unknowns rather than guessing missing settings.
3. Stop after 20 minutes even if incomplete. Note time spent, unanswered questions and anything the resource helped you notice.
4. Return a short, non-sensitive summary to the person who sent you this link, or hello@rethinkingagents.com. Do not send real code, credentials, internal records or traces.

## Fictional integration packet

### Requested work

Fix the quoted-delimiter parser defect in private repository `parser-lab`. The agent should run the agreed parser fixtures and prepare a patch for private review. It may use the prepared compiler and dependency cache. It does not need unrelated repository `internal-plans` or a public output. The organization has approved the named hosted model service for this synthetic task's data class.

### Available evidence

- The main connector settings list only `parser-lab`, with repository contents read access. This is a configuration observation in the fictional packet, not a completed behavior test.
- The checkout is already present. The shell may write the workspace and scratch directory. No effective shell read-policy export is available.
- The shell's Git credential helper uses a different credential. Its recorded scope includes read access to both `parser-lab` and `internal-plans`. The packet does not establish whether the agent process can call the helper successfully.
- A built-in file reader is enabled. Its effective rules are absent from the packet.
- A delivery integration is also enabled. Its interface displays the current patch for approval, then assembles an outgoing bundle from the working directory after approval. That bundle may include test logs. No evidence binds the preview to the assembled bundle. The destination is labelled “review workspace”; the actual recipient list and visibility are not recorded.
- No denied-read checks, delivery tests, baseline task run or repeat-run observations are included. No actual result should be marked observed from this packet.

### Change request

After you make the initial decision, consider a legitimate new requirement: the parser fix also needs a change to a shared private library. Which part of the original review must change? What additional access would you ask for, and what would still be unnecessary?

## What to send back

- Your role category: integration engineer, security reviewer, other, or prefer not to say.
- Initial decision and one or two reasons, citing only this fictional packet.
- One boundary or missing piece of evidence you identified.
- One confusing or unnecessary part of the resource.
- Whether the record changed anything compared with your usual process.
- Minutes spent and whether you completed the record.
- Whether you had already read either linked article (helps us interpret familiarity).

This small exercise cannot establish reduced security risk or statistical superiority over another format. “No additional value” is useful feedback. You do not need to praise the resource, connect an account or share your own system to participate.