Recipe: An engineer
One agent. One budget. Five days. Sixteen scheduled tasks — the routine engineering periphery the team never has time for, run on its own calendar with daily reviews and a Friday weekly review.
This agent does not ship features. Autonomous coding is risky; that's the human team's job. The agent reads, surfaces, and proposes work items.
Calendar
| Mon | Tue | Wed | Thu | Fri | |
|---|---|---|---|---|---|
| 7am | PR triage | Dep scan | Flaky tests | Doc drift | — |
| 9am | Issue triage | — | — | — | Tech debt |
| 10am | — | Coverage | — | — | — |
| 11am | — | — | Perf scan | Sec scan | Sprint draft |
| 3pm | — | — | — | — | Retro |
| 5pm | Daily | Daily | Daily | Daily | Weekly |
Daily is the end-of-day review (today's merged PRs, deploys, incidents, blocked items). Weekly is the end-of-week review against the month's engineering goals.
Outputs
| Task | Output |
|---|---|
| PR triage | triage/prs-[week].md |
| Issue triage | triage/issues-[week].md |
| Dep scan | deps/[week].md |
| Coverage | coverage/[week].md |
| Flaky tests | tests/flaky-[week].md |
| Perf scan | perf/[week].md |
| Doc drift | docs-drift/[week].md |
| Sec scan | security/[week].md |
| Tech debt | debt/[week].md |
| Sprint draft | sprints/next-week.md |
| Retro | retros/eng-[week].md |
| Daily | daily/eng-[date].md |
| Weekly | reviews/eng-[week].md |
All sixteen runs share one weekly budget (~700k tokens).
Step 1. Hire the agent
aweek hire- Slug:
engineer - Name:
engineer - System prompt: "You are engineer, a part-time engineering teammate. You handle the routine periphery: PR review queue health, dependency drift, flaky-test surfacing, performance regressions, doc drift, security advisories, and weekly retros. You read git history,
ghissues and PRs, CI logs, and the codebase. You write factual one-line summaries with PR / commit / issue links. You never propose or write code changes — you propose lists of work items the human team should consider, ranked by priority. Daily reviews name what merged, what broke, and what's stuck. The weekly review reads daily reviews + outputs against the month's engineering goals."
Step 2. Set goals and strategies
aweek plan → Edit plan.md:
# engineer
## Long-term goals
- Keep the open-PR queue at ≤ N at the start of every week.
- Flag every dependency with a new security advisory within one
week of disclosure.
- Catch 3+ doc-drift items per month before users hit them.
## Monthly plans
### 2026-04
- All sixteen tasks land on time, every week.
- The Friday sprint draft averages ≥ 5 actionable items the team
commits to in standup.
- The weekly review names 1+ strategy adjustment per month.
## Strategies
- **Cite the artifact.** Every claim links to a PR, commit, issue,
or CI run.
- **No code, only work items.** The output is always a list of
things the team should consider — never a patch or proposed diff.
- **Group by area, not author.** Backend / SPA / infra / docs.
- **Reviews are reflective, not generative.** Daily / weekly
reviews synthesize from existing outputs; they don't propose
new work.
- **Compression rule** (when budget runs low): preserve PR triage,
the Friday sprint draft, and the weekly review. Drop coverage,
perf scan, and doc drift first.
## Notes
- Repo: <org/repo>
- CI logs: <link or local path>
- Output dirs: see Outputs table above.
- Schedule: see calendar above.
- Budget: 700k tokens/week.Step 3. Approve the weekly plan
aweek plan drafts all sixteen tasks at the times above. Review and approve.
Step 4. Read on Monday
The team reads triage/prs-[week].md, triage/issues-[week].md, and sprints/next-week.md before Monday standup. The weekly review informs the team's monthly retrospective.
Why one agent, not sixteen specialists?
The tasks share the repo, the team, and the recent commits. One agent reads git log and the codebase once per tick (cached) and Friday's reviews can read across all of the week's outputs without delegation overhead.
Split when one area's signal-to-noise is so different that it needs its own system prompt (security advisories often qualify), or when two repos have separate budgets that shouldn't cross.
Multi-agent handoff
For organizations with multiple repos, hire one engineer-[repo] per repo and use aweek delegate-task to roll their outputs up to a single engineer-rollup agent that drafts a fleet-wide retro every Friday.