The Signal FilesInteractive Field-GuideCompanion to the research

The harness, made legible.

A capable agent in a bare room still fails. This is the companion field-guide to our Signal File on harness engineering, the discipline that turns a powerful model into a reliable one. Here it is mapped against the Interpretable Context Methodology (ICM). Twelve lectures and one paper, made navigable.

12
Lectures
6
Projects
5
Subsystems
9
Agreements
7
Divergences
Provenance

We didn't make the training. We made it legible. The curriculum is Learn Harness Engineering by walkinglabs (MIT-licensed); the interpretable-context method is the work of Van Clief & McDermott (arXiv:2603.16021). Stop Trying To Be Invisible did not write the course or the paper. We built this visual structure to turn dense source material into something you can navigate and use in an afternoon. The thinking is theirs; the legibility is ours. Read the full research →

01 · THE MENTAL MODEL

A harness is not a prompt. It's a working loop.

The course's core claim: a harness doesn't make the model smarter. It builds a closed-loop system around the model so its capability actually gets realized. Select a phase to see what it does and which file carries it.

Learn Harness Engineering, course homepage harness loop (AGENTS.md → init.sh → run → feedback → verify → handoff)
02 · THE CURRICULUM

Twelve failures, twelve fixes.

Each lecture diagnoses one way capable agents fail, then introduces the harness primitive that prevents it. Select any card for the thesis, the failure, the fix, and the artifact you'd add to a repo.

03 · THE PRACTICE

Six projects, one app, an ever-stronger harness.

The lectures are theory; the projects build it. One knowledge-base app is carried through six stages. Each adds a layer of the harness and is run twice, weak versus strong, so you feel the difference rather than read it.

04 · THE CORRELATION

Harness Engineering × ICM

Two methods, one intellectual lineage: Unix, separation of concerns, plain text, the repository as truth. But they optimise for different things. Here's exactly where they agree, where they part, and how their concepts line up.

HARNESS ENGINEERING

Reliability for autonomous coding agents

Make Codex / Claude Code build software dependably via five subsystems (instructions, tools, environment, state, feedback), with executable verification replacing the agent's "feelings." Optimises the agent's reliability surface.

ICM · VAN CLIEF & McDERMOTT, 2026

Interpretability for human-reviewed pipelines

Replace framework orchestration with folder structure: numbered stages, markdown context, a human review gate at every boundary. Optimises the human's control & interpretability surface.

ICM context layer / moveHarness Engineering equivalent
The one-line synthesis

ICM is the macro skeleton: where files live, what context loads when, and where a human reviews. Harness Engineering is the micro machinery inside one stage: how you prove the output is real, how state survives a restart, how the run stays observable. They aren't rivals. You can run a harness-engineered coding stage inside an ICM pipeline: ICM hands the agent a clean, scoped context; the harness makes sure what comes back is true.

05 · THE CHEAT SHEET

Failure → fix → the one file that solves it.

The operating principle: add the smallest artifact that addresses the failure you actually see. Don't solve every reliability problem by dumping more text into one global instruction file. Orange marks the broken thing.

Failure modeWhat it looks likePrimary fixArtifact
Learn Harness Engineering, resources/reference/method-map
Harness strength · the same model, a different outcome
InvisibleEmergingVisibleDominant
P1 Prompt only: the agent guesses, drifts, declares victory early
P3 +State & continuity: survives restarts, one feature at a time
P5 +Role separation: generator / evaluator / planner
P6 Complete harness: observable, verified, clean handoff

The variable that moves results isn't the model. It's the harness.

Role separation alone takes an evaluator score from 1.6 / 5 (one agent grading itself) to 4.9 / 5 with a planner, generator, and independent evaluator. That gap is the whole discipline in one number.

Learn Harness Engineering, Project 05, evaluator-rubric scores (single-role → plan-gen-eval)
06 · START TODAY

The minimal harness you can ship this afternoon.

You don't need all twelve lectures to begin. Drop four files into any repo and you've crossed the line from prompting to a harness.

FILE A

AGENTS.md

A 50–200 line router: overview, run commands, ≤15 hard constraints, links out. Not an encyclopedia.

FILE B

init.sh

One script that takes a fresh session from clone to a passing example test in minutes.

FILE C

feature_list.json

Each feature as a triple: behaviour · verification command · state. The machine-readable source of "done".

FILE D

claude-progress.md

Clock-out file: what's done, what's broken, what's next. The next session reads this first.

07 · SOURCES

Where this comes from.

  1. The Signal File, our research piece (start here)
  2. Learn Harness Engineering, course (MIT-licensed; twelve lectures, six projects)
  3. Interpretable Context Methodology, Van Clief & McDermott, arXiv 2026 (CC BY 4.0)
  4. OpenAI: Harness engineering, Lopopolo, 2026
  5. Anthropic: Effective harnesses for long-running agents, Young, 2025
  6. Anthropic: Harness design for long-running application development, Rajasekaran, 2026