Field notes from
production self-improvement.

How Reflex is built, and what we see when we let it run on real benchmarks.

Field guideApr 12, 2026 · 6 min

How I run 20 Claude Code agents in parallel without git conflicts.

Agents could run in parallel, but our tooling forced them one at a time. Git worktrees give every agent its own branch and directory, and Workstreams, the open-source IDE we built around it, orchestrates the rest.

Read post →
Run reportMay 3, 2026 · 9 min

What happens when you let a self-improving harness run unattended on telecom.

A single Reflex run on tau-bench telecom. Sixteen iterations, no human prompt engineering. The agent moved from 0.400 to 0.875 verified accuracy and wrote ten of its own protocol rules.

Read post →