Outpost · 90-day embedded engineering: Launch one AI agent in 90 days.

1. 1

   Weeks 1–2

   Agree what safe enough means

   Map what the agent can do, where mistakes matter, who approves important actions, and what must be true before launch.

   You keep

   A written release plan with clear owners and pass-or-fail tests.

2. 2

   Weeks 3–6

   Test it and add controls

   Build repeatable tests, limit access to tools and data, add human approval where needed, and practise failure and recovery paths.

   You keep

   Automated checks, safer permissions, fallbacks, and a record reviewers can follow.

3. 3

   Launch and 90 days

   Release, watch, and hand over

   Support the launch, track real failures, turn each lesson into a permanent test, and transfer ownership to your team.

   You keep

   A live operating routine and a system your team can continue without us.

Throughout, your team gets one clear place to see what was tested, what is controlled, who approved the launch, and what still needs attention.