Before writing code, shadow the manual process for 2-3 days. Document every input the human uses, every decision they make, every system they touch, and every output they produce. Include edge cases — what happens when data is missing, when the customer is angry, when the policy is ambiguous.

Then cut ruthlessly. Your MVP covers one happy path that delivers 80% of the value. If the full workflow has 12 steps, your MVP handles the 5 most common. Edge cases get manual fallback, not automated handling.

Deliverable: a one-page workflow diagram with inputs, agent steps, tool calls, human checkpoints, and outputs. Share it with your design partner and ask: 'If the agent handled these 5 steps reliably, would you pay for a pilot?' Their answer determines whether you build.

Week 2: authentication, one core integration (the system your agent reads from most), basic agent loop (prompt + 2-3 tools), and a minimal UI that shows input and output. No polish — function over form.

Week 3: add RAG or additional tools, implement the human approval step, and dogfood internally with 10 real examples from your design partner. Target 'works correctly for 10 real examples' not 'handles every possible input.' Measure completion rate on those 10 examples daily.

Resist feature creep. The most common Week 3 mistake is adding a second workflow before the first one works. If your ticket triage agent handles 7/10 examples correctly, fix those 3 failures before adding email drafting. Completion on one workflow beats partial coverage of three.

Launch a paid pilot with one design partner at $500-2,000/month. Paid is critical — free pilots produce feedback; paid pilots produce commitment. Set expectations: weekly 30-minute calls, fixed SLA (agent handles X tasks per day), and manual fallback when confidence is low.

Track three numbers daily: tasks attempted, tasks completed successfully, tasks escalated to human. Share these with your design partner every week. Transparency about what works and what does not builds trust faster than overpromising.

The goal of Week 4 is learning, not scaling. You want to discover what breaks in production with real data, real users, and real stakes. Every failure mode you find now saves weeks of rework later. One paying pilot customer with honest feedback is worth more than 100 free signups.