CloudThinker Engineering

Home > Journeys > Agent-native team

Agent-native team

The question: How should a small team work when agents do most of the building, and what do people still need to do?

In October 2026 a team of about five engineers merged 2,901 merge requests, most of them written, tested and merged by agents. This journey is about how we work with agents at that rate, and what broke on the way.

This page sums up the journey so far. The full write-up is the entry below. Start reading: Entry 1: Lessons learned from building a software factory (6 min).

Entries (the full posts)

  1. Lessons learned from building a software factory 2026-11-08, start, 6 min, 1.6k tokens NEW
    How a team of about five engineers went from about 40 to 2,901 merges a month: talk before code, one control folder, proof instead of review, small merge requests and a shared test cache.

What we think now

Revision 1, 2026-11-08.

  1. Talk first, code last. A discussion with an agent ends in one HTML decision page; the team agrees on it, it becomes tickets, and only then do agents write code.
  2. Every agent session starts in one control folder of markdown notes, skills and machine guides, so each session starts with what the last one learned.
  3. Agents merge their own small changes. Proof replaces the human reviewer. Every merge request starts the full app, captures screenshots and records a video. No proof, no merge.
  4. Small merge requests, not smarter agents, made auto-merge safe. In October 2026, 2 of 2,901 merges were reverts.
  5. The first limit was the machine, not the model. A shared test cache and a shared check runner cut the median wait for checks from 302 to 73 seconds.

Open questions


Follow this journey: RSS. Plain text version: /journeys/agent-native-team.md


CloudThinker Engineering · engineering.cloudthinker.io · Last built 2026-11-08 · Plain text version