Why agent change windows belong in your production calendar
Prompt edits and new tools are production changes. Publish when agents may change, who approves, and how you roll back, the same way you schedule app deploys.
This week's headlines roundup highlights agent change windows beside normal release calendars. That is overdue. Most teams still let anyone with console access push a prompt at 4:45 PM on Friday, then wonder why Monday support volume spikes.
If you already run injection red teams and eval gates, change windows are the operational wrapper: they tell humans when those controls must pass before traffic sees a new version. The Jan 22 attestation headlines set the procurement bar this builds on.
Define what counts as a change
Include anything that can alter customer-visible behavior: system prompts, tool lists, model IDs, retrieval indexes, rate limits, and feature flags that enable write tools. Exclude pure observability tweaks only if they cannot change outputs.
Publish a simple weekly rhythm
- Change window: e.g. Tue and Thu 10:00 to 14:00 local, staging first.
- Freeze: no prod agent changes outside the window except incident break-glass.
- Owner: named approver per workflow (support lead + engineering).
- Rollback: previous prompt and tool manifest kept one click away.
Wire promotions through pipelines, not console edits, so the calendar is enforceable, not aspirational.
Tell customer-facing teams before you ship
Send support and account managers a one-line release note when an agent changes: what is different, what to escalate, and whether SLAs shift. That prevents reps from defending behavior you already retired.
Pair changes with monitoring
During the window, watch error rate, handoff rate, average turns, and tool failure rate for an hour after promotion. If metrics move outside bands, roll back inside the same window and open an incident if customers were affected. Tie thresholds to published SLAs and re-run injection drills after any prompt or tool change.
Break-glass without chaos
Incidents still need hotfixes. Log every break-glass change with ticket ID, approver, and post-incident review. Restrict break-glass roles with IAM-aligned permissions so emergency access does not become the default path.
Week-one rollout
- Pick your highest-traffic customer agent.
- Document current prompt and tool manifest as the rollback baseline.
- Announce one recurring change window to engineering and support.
- Move the next planned prompt edit into that slot and measure support impact.
Subprocessor DPAs and mTLS from this week's headlines protect data in transit and in contracts. Change windows protect customers from your own shipping velocity.
Need help operationalizing agents on your site? Talk with us or see the Website Agent service.
Matt Potter · Swift Media