Learn Harness Engineering
A project-based course for building the environment, state management, verification, and control mechanisms that make AI coding agents work reliably.
- Source repo
- walkinglabs/learn-harness-engineering
- Stars
- ★ 21k
- Last updated
- 18d ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 32/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Engineers using Claude Code or Codex who want to reduce agent failures like skipping steps, breaking tests, or prematurely declaring victory.
- Not a fit if
- Readers who want theory only and will not do the projects
- Users needing detailed Codex- or Claude Code-specific configuration
- Source review
- 32/100 · Major gaps 8 safety controls not found
What does this agent do, and when should you use it?
Learn Harness Engineering is a project-based course dedicated to the engineering of AI coding agents. It synthesizes the most advanced harness engineering theories and practices from industry leaders like OpenAI and Anthropic, and integrates resources from the Awesome Harness Engineering list. The course comprises 14 lectures and 8 hands-on projects, all centered around building an Electron-based personal knowledge base desktop app, evolving from a prompt-only baseline to a complete production-grade harness. The core idea is constructing a complete working environment around the model, consisting of five subsystems: instructions, state, verification, scope, and session lifecycle, implemented through files like AGENTS.md, init.sh, feature_list.json, and claude-progress.md. The course is available in 15 languages, includes a PDF build pipeline, a VitePress documentation site, a reusable skill for scaffolding harnesses, and a dependency-free audit script. It targets engineers, researchers, and tech leads who want to improve the stability and quality of AI coding agents.
This repository is an educational course that teaches harness engineering for AI coding agents. It provides 14 lectures (L01-L14) covering topics from foundational concepts to loop and graph engineering, and 8 projects (P01-P08) that guide users through building a complete harness for an Electron app. The course delivers concrete templates and scripts: AGENTS.md and CLAUDE.md for agent instructions, init.sh for environment initialization and health checks, feature_list.json for tracking feature completion, and claude-progress.md for session progress. It also includes a harness-creator skill that scaffolds production-grade harnesses, and an audit-harness.sh shell script that checks a repository against the five harness subsystems. The content is delivered via a VitePress documentation site and a resource library with templates in 14 languages. Users can run npm run docs:dev to preview locally, and npm run pdf:build to generate PDF coursebooks.
- Engineers using Claude Code or Codex who want to reduce agent failures like skipping steps, breaking tests, or prematurely declaring victory.
- Tech leads who need to understand how environment design affects agent performance and want to standardize harness practices across their team.
- Researchers seeking a systematic understanding of harness design patterns and trade-offs, with references to industry papers and blog posts.
- Learners who want to follow a hands-on curriculum, building a harness incrementally through 8 projects and comparing before/after results.
- Developers who want to quickly scaffold a production-grade harness for their own project using the harness-creator skill.
- Teams wanting to audit existing repositories for harness best practices using the dependency-free audit-harness.sh script.
How do you install or deploy this agent?
git clone https://github.com/walkinglabs/learn-harness-engineering.git
cd learn-harness-engineering
npm installHow do you use this agent?
To preview the documentation locally, run npm run docs:dev and open the URL printed by VitePress. To build PDF coursebooks, run npm run pdf:build; output is written to artifacts/pdfs/. To use the harness-creator skill, follow the instructions in skills/harness-creator/ to install it into your IDE or agent workspace. To audit an existing repository, run bash tools/audit-harness.sh /path/to/your/repo.
What are this agent's strengths and limitations?
- Project-based learning anchored on a real Electron app, making abstract concepts tangible and providing measurable before/after comparisons.
- Covers the five core harness subsystems with actionable templates (AGENTS.md, init.sh, feature_list.json) and a session lifecycle diagram.
- Translated into 15 languages with a PDF build pipeline, broadening accessibility for non-English speakers.
- Includes a reusable skill (harness-creator) and a zero-dependency audit script (audit-harness.sh) for immediate application to user projects.
- Requires familiarity with terminal, git, and at least one programming language; not for complete beginners.
- Assumes access to a coding agent tool like Claude Code or Codex; without one, the projects cannot be completed as intended.
- The example app is Electron-specific, which may not resonate with all developers, though the principles are transferable.
- It does not provide detailed tool-specific configurations for Codex or Claude Code, leaving some adaptation to the user.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Learn Harness Engineering This agent | 32 · Major gaps | CLIFree + model costs | ★ 21k | 18d ago | TypeScript | Codex · Claude Code |
| Harness Engineering Series | 26 · Major gaps | Web appFree | ★ 3.2k | 6mo ago | Python | Codex · Claude Code |
| Better Harness | 85 · Good | Agent plugin / skillFree | ★ 2.4k | 6d ago | JavaScript | Codex · Claude Code |
| Munder Difflin – Multi-Agent Office Floor | 43 · Major gaps | Desktop appFree + model costs | ★ 8.9k | today | TypeScript | ChatGPT · Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows: the repository is a teaching course, containing no executable code or agent logic, so there are no permission requests, data flows, or external side effects. No user confirmation mechanisms, sensitive data handling, or rollback mechanisms. Dependencies are only for documentation building, no security audit found. Source attribution: README cites OpenAI and Anthropic articles but without specific versions or verification. Deductions: since there is no actual agent behavior, all trust criteria score 0; source attribution only partially satisfied because citations are unverified.
Evidence shows: README and course structure are consistent in describing course content, no obvious contradictions. Dependencies (vitepress, playwright, etc.) are common and available, but no version pinning or integrity checks. Failure messages: no error handling or user prompts provided. Deductions: self-consistency and dependency availability score 2 because documentation structure is clear and dependencies are common; failure messages score 0 because no relevant evidence.
Evidence shows: course targets beginners, provides multi-language support and phased learning paths, adapting to different scenarios. Capability boundaries: course clearly states it is for teaching, but does not detail its boundaries as an agent product. Trigger precision: no trigger conditions defined. Environment fit: provides installation and build instructions, but does not specify compatibility across OS or environments. Deductions: audience and scenarios score 2 for multi-language and phased design; capability boundaries score 1 for unclear boundary description; trigger precision scores 0 for no trigger mechanism; environment fit scores 2 for build instructions.
Evidence shows: information architecture is clear with TOC and chapter structure. Install notes: npm scripts and build commands provided. Naming stability: filenames like claude-progress.md are retained, but naming conventions not explained. Examples and FAQ: project examples provided, but no FAQ. Known limitations: not mentioned. License: MIT license explicit. Versioning and changelog: version 0.1.0 exists, but no changelog. Maintenance responsibility: not clearly stated. Deductions: information architecture, install notes, examples, and license score 2; naming stability scores 1 for unclear conventions; known limitations scores 0 for absence; versioning and changelog scores 1 for only version number; maintenance responsibility scores 1 for not explicit.
Evidence shows: output usability: course provides templates and resources directly usable. Marginal value: course offers unique perspective, but as an agent product, direct value is limited. Cost-benefit: course is free but requires time investment. Deductions: output usability scores 2 for practical templates; marginal value scores 2 for unique content; cost-benefit scores 1 for no cost analysis.
Evidence shows: claim traceability: README cites OpenAI and Anthropic articles but without specific links or versions. Cross-source corroboration: multiple sources cited but no cross-validation. Fact-inference separation: course content distinguishes facts and inferences but not explicitly labeled. Deductions: claim traceability scores 1 for incomplete citations; cross-source corroboration scores 1 for multiple sources but unverified; fact-inference separation scores 1 for not explicit labeling.
- Not found in source: least-privilege scopingGrant only what the task needs: a dedicated account or read-only token, scoped to specific directories and repos.
- Not found in source: confirmation before actingTurn on (or add) a confirmation step before it acts, and try it in a sandbox or test environment before real data.
- Not found in source: data-flow disclosureWatch which external services it contacts (proxy or firewall logs) and keep sensitive data out until you know where it goes.
- Not found in source: sensitive-data handlingUse dedicated, low-privilege, revocable API keys — never production credentials — and keep secrets out of logs.
- Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
- Not found in source: disclosed external effectsEstablish which external systems it writes to, sends to or changes, and verify with test accounts or repos before production.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- Not found in source: verifiable attributionInstall from the official repo or registry and check the publisher and URL to avoid look-alike packages.
- This repository is a teaching course, not an executable agent product, so trust-related criteria are not applicable.
- Dependencies are not version-pinned, posing supply chain risk.
- No known limitations or failure handling documentation provided.