Part 1 · Day 1 — Nov 30, 2026

DON’T TRAIN THE MODEL, EVOLVE THE HARNESS

When an agent fails, most teams open the prompt and start editing. This workshop hands that job to another agent. A meta-harness reads failed trajectories, writes a patch to the harness around a fixed model, and keeps the patch only if it wins on a validation split. In one run it took GPT-OSS-20B inside OpenCode from 4.8% to 14.8% on held-out Terminal-Bench 2.1 for $50 of API credits. The weights stayed the same. Small fixes carried most of the gain. One made the agent test its code before ending the session. Another repaired a tool call that OpenCode rejected over a single stray bracket. We’ll also look briefly at a legal agent benchmark, where the same loop raised DeepSeek-V4-Pro from 63% to 80% of grading criteria.

Then you run your own loop. The target is a small coding agent, 150 lines of Python with 20 unit-tested tasks. A cheap model does the work and Claude Sonnet proposes the patches. Each iteration takes a minute or two, so you’ll watch several patches land and read the diffs as they arrive. While your loop runs, we walk through the real Terminal-Bench run patch by patch. You leave with the reference code and an onboarding guide that turns your own agent and eval into a spec you can run at home.

For engineers who build agents and have an eval for them, or want one. Python and the command line assumed.

Logistics: about 15 minutes of setup before the session from a prep email. The loop runs on the laptop with no Docker and no GPU. Model inference and the proposer run on sponsor credits.

Part 2 · Day 2 — Dec 1, 2026

YOUR BENCHMARK SCORE IS A HARNESS SCORE

A model’s benchmark number depends on the harness that ran it. On 250 held-out SWE-bench Pro tasks, swapping the harness around GLM-5.2 moved pass@1 from 23% to 52%. The ranking of ten harnesses on one model told you nothing about their ranking on another. Cost moves as well. Running Muse-Glimmer-30B on Terminal-Bench 2.1, OpenCode solved tasks at $1.73 each and OpenClaw at $8.80, with identical weights. Most of that gap came from how often each harness re-sent its context.

In this workshop the room runs the experiment. Each table takes two harnesses and runs Muse-Glimmer-30B on a 10-task slice of Terminal-Bench 2.1 in hosted sandboxes. You pull pass rate, tokens and cost from your own logs and add them to a shared leaderboard. The room’s Pareto plot builds up on screen next to a reference set of 1,500 attempts. While jobs run, we open failed trajectories on the projector and work out why each harness lost. You leave knowing how to pick a harness for your own model and budget, and how many tasks and trials you need before a gap between two harnesses means anything.

For engineers who ship coding agents or choose models for them. Python and the command line assumed.