Coffee first. Then five checks. Each one takes a few minutes, and each one lines up with something that changed for agent operators this week.
1. Grep for model IDs that shut off
Some older models went dark this morning. An agent still pointing at one won't slow down. It fails on its first call.
- Search your configs, env files, and prompts for every hard-coded model ID.
- Check each against your provider's deprecation page.
- Swap and pin a replacement before the scheduled jobs start, not after the first page.
2. Put a ceiling on retries after a refusal
Some refusals that used to be free now cost money. A retry loop that re-asks the same thing five times now pays five times.
- Treat a refusal as a final answer, not a transient error.
- Cap retries per task, and don't retry on refusal at all unless the input changed.
- Look at yesterday's logs: how many refusals got retried?
3. Check that dashboards fail loud on empty fields
An audit feed stopped returning file names, even on old records. Anything keyed on that field goes blank without an error.
- For each audit or event feed you depend on, list the fields your dashboards and alerts key on.
- Add a check that alerts when a field you rely on comes back empty or missing, rather than drawing an empty chart.
4. Bump pinned CI action versions
A CI runtime lost its last opt-out. Agent CI jobs and PR bots on old action versions can break.
- Find the workflows your agents and PR bots use.
- Update pinned action versions and run each once by hand today.
5. Test new models on a copy of real traffic
One new model this week is cheap enough to use for subagents. Another rejects request shapes that worked last week.
- Replay a small sample of real requests before switching anything.
- Watch for schema or parameter rejections, not just output quality.
- Move subagents to the cheaper model only after the replay is clean.
Copy/paste
[ ] Model IDs grepped, pinned, none on today's shutdown list
[ ] Retries capped; no retry on refusal
[ ] Alerts fire on empty/missing audit fields
[ ] CI action versions bumped; agent workflows run once by hand
[ ] New models replayed on real requests before rollout
If a box stays unchecked, hold that part of the rollout until it's done.
These come from this week's Weekstart Brief, a Monday rundown of what changed for people running agents. The free teaser is here: https://stackyard.fyi/issue-003-teaser
Top comments (0)