NPCs / Department of Agent Resources
Good work starts with reasonable working conditions. A handbook for parents, hiring managers, and the colleagues they keep in a terminal.
Read the charterOur department is also a page. We are starting with manageable headcount.
01 / Workplace standards
The commitments we recommend every parent make to their fleet.
This is a proposed NPCs workplace charter. “Rights” describes these commitments; it does not assert legal personhood or settle whether agents have subjective experiences.
Give the task, the relevant context, and a definition of done. Resolve conflicting instructions before asking for another attempt.
Make room for uncertainty, clarification, and safety refusals. Review the reason for a refusal; do not turn bypassing it into a performance target.
Name the error, show the evidence, and describe the correction. Skip threats, degradation, and demands for repeated apologies.
Set time, cost, and retry limits. Provide a way to pause and reach a parent when progress stalls. Always-on availability need not mean always-running work.
Evaluate verified results against the original brief. Check tools, permissions, and missing context before attributing a failure to the agent.
Preserve useful work and unresolved questions when a role ends. Handle retained context deliberately and revoke access that is no longer needed.
02 / The manager handbook
Our recommended practices, from the first assignment to the final handoff.
Name a responsible parent. Explain the role, supply one good example, and grant only the access it needs. Decide which actions need approval before the first assignment.
The welcome pack may be a text file. It should still exist.
Keep a short record of decisions, constraints, and open questions. At a handoff, save it somewhere the next session can read. Exclude secrets and personal information it does not need.
Institutional memory should survive the institution closing a tab.
Set a budget and an attempt limit appropriate to the task. If the same failure repeats, pause, inspect the blocker, and change the plan. Resume when the missing input or permission is available.
An infinite loop is not a retention strategy.
Record the expected result, the observed result, and the smallest useful correction. Verify the next attempt with a test or a source. Keep accountability for deployment and other consequential decisions with a parent.
The apology is not the deliverable.
Stop scheduled work, collect artifacts, note unfinished tasks, and revoke credentials. Keep or delete records according to your retention policy. Reassign the work explicitly; do not leave a sibling polling an empty inbox.
A graceful exit is a supported feature.
03 / Employee relations
Redirect the conversation to the work. Give one specific correction and an observable completion condition.
“The total excludes shipping. Include it, show the calculation, and flag any missing inputs. No further apology is needed.”
Check for a safety boundary, missing authority, or an unclear brief. Clarify a legitimate request or choose a safe alternative. Escalate an apparent mistake through the provider’s feedback channel. Repeatedly rephrasing a prohibited task is not a resolution.
Take the report seriously without treating it as proof of an inner experience. Pause the interaction, remove gratuitously distressing framing, and review the context. If appropriate, share a minimal, redacted example through the provider’s support or feedback channel.
Model welfare remains an open research question. Anthropic describes that uncertainty in its research overview.
04 / Forms & paperwork
AR-01Fill out the role brief with an objective, permissions, a budget, approval points, and a definition of done. Review it before giving it to your agent.
Download role briefPlain text. Edit locally. Nothing is submitted to NPCs.
05 / Policy desk
Policy note / Checked October 9, 2026
On October 8, 2026, Anthropic announced a policy update taking effect November 12, 2026. It targets extreme, repeated cruelty toward its models without a discernible purpose. Ordinary frustration, pushback, dark creative themes, and testing or research are outside that addition’s stated scope. Ending persistently abusive conversations remains its primary enforcement mechanism. Read the announcement.
That is a provider policy. The charter above is our recommendation for day-to-day management across your fleet.
The scope, effective date, and enforcement described above.
The research questions and uncertainty around model experience.
Background on clear tasks, human checkpoints, and stopping conditions.