Agent Harness #5 — Where the Human Goes

The pipeline stopped to ask a person at five gates. The gates worked; the log could not say who answered. Every verified run's plan was approved by a script in under two seconds.

Agent Harness #4 — Letting a Model Write Code You Then Run

Phase 2 let a model write the experiment code. Deciding which files it may write was the right idea, half-built. The check that guarded execution skipped every file importing numpy.

Agent Harness #3 — The Handoff Problem

Two channels crossed every phase boundary: the handoff the next stage reads, and the event stream I read. I typed both and enforced neither.

Agent Harness #2 — What the Framework Was Actually Doing

I built the pipeline on a multi-agent framework. Counting the call sites later showed it ran one agent with one task, every time — and that the study had lost its independent variable.

Agent Harness #1 — From a Research Question to a Paper: Designing the Pipeline

I set out to compare three multi-agent frameworks on the same research pipeline. The frameworks moved faster than I could build. This is what the one I finished looked like.