Your 30-Day Agent Rollout

Put it together. Thirty days is enough to get one agent running reliably if you keep the scope small.

The Plan, With Deliverables

• Days 1 to 3: choose the use case and write the design page. Deliverable: scored table and design page.
• Days 4 to 7: write the skill and the limits, and build the six-input test table. Deliverable: SKILL.md and a table with all six inputs run.
• Days 8 to 12: connect read-only access and run stage one. Deliverable: an agreement rate on 30 real or synthetic items.
• Days 13 to 30: add drafts, review every output and record one results row each week (the measurement exercise). Two or three rows land by day 30. Deliverable: an error log with fixes, a rerun of the test table and your weekly rows.
• Days 27 to 30: write the run guide, set a weekly check and plan the second agent. Deliverable: the run guide and a dated calendar.
• After day 30: the fourth weekly row lands around day 40. Make the keep, fix or stop decision then. It comes after four weekly rows, so it lands after day 30.

Gates

• End of day 7: all six test inputs pass, including the injection input.
• End of day 12: agreement of at least 90 percent, or fix the skill before adding drafts.
• After four weekly rows: net time saved after review, or stop.

The Run Guide (Copy This)

• What the agent does, in two sentences.
• What it can access and what it can't.
• Its limits and stop rules.
• How to check its work in under a minute.
• How to switch it off, step by step.
• Who is responsible and who to tell if it fails.
• The date of the next review.

Common Mistakes

• Skipping the gates because it feels like it's working. Fix: write the pass numbers in advance.
• Letting scope grow in week two. Fix: put new ideas on the second-agent list.
• No named owner. Fix: write a name in the run guide.

Exercise

• Copy the plan into your calendar with real dates, including the date of your fourth weekly row, around day 40.
• Write the run guide now, even in draft.
• Pass condition: dated deliverables on your calendar and a run guide that someone else could use to pause the agent.