harness-retro

Installation
SKILL.md

Harness Retro

Look at what happened in an agent run and propose changes to the harness so the next run goes better. You propose. The user decides. Do not edit the harness until the user picks which findings to apply.

An agent is a model plus a harness. The model is the same for everyone. The harness is the part the team owns, and it is the part that can get better. Every run teaches the agent something: how to reach the database, which test is flaky, which command fails the first time. That lesson dies with the session unless someone writes it into the repo. This skill is that someone.

Before you decide where any fix should live, read references/harness-primer.md. It is a short summary of the harness layers and the rules for picking one, taken from the book AI-Native Software Engineering.

Step 1: Gather the evidence

Read these sources, in this order:

  1. The current conversation. It is already in your context. Walk it from start to end.
  2. Sub-agent reports. Any results that sub-agents returned during the run. Read their "odd things" notes closely: the lead agent only sees these summaries, so a problem a sub-agent hit and fixed quietly is easy to miss.
  3. Past sessions in the same repo. A problem seen in three runs is a much stronger signal than one seen once. Read the transcripts or logs the user points you to. If they point to none, and your agent tool keeps session history you can read, look at the last few sessions for this repo. Skim for user corrections, failed commands and steps repeated by hand; don't read them line by line. If there are none, carry on and say so in the report.

These are signals:

Installs
20
GitHub Stars
16
First Seen
Oct 26, 2026