Learn Harness Engineering

A project-based course by Walking Labs on the engineering of AI coding agents, paired with the awesome-harness-engineering curated list. It synthesizes the OpenAI and Anthropic harness writeups into a practical curriculum and gives the wiki a source-backed curriculum for Harness Engineering. Source: X/@_vmlops, 2026-05-17

What it teaches

A harness does not make the model smarter; it establishes a closed-loop working system around the model. The course teaches how to make tools like Codex and Claude Code reliable by constraining them with explicit rules and boundaries. Core skills: Source: Walking Labs, learn-harness-engineering, 2026-05-17

• Constrain agent behavior with explicit rules and boundaries.
• Maintain context across long-running, multi-session tasks.
• Stop agents from declaring victory too early (the verification gap).
• Verify work using full-pipeline tests and self-reflection.
• Make the runtime observable and debuggable.

Structure

The course splits into theoretical lectures (starting with "Why Capable Agents Still Fail"), hands-on projects (for example "Baseline vs Minimal Harness"), and a copy-ready resource library. The templates ship a minimal harness pack: AGENTS.md, feature_list.json, and claude-progress.md. Source: Walking Labs, learn-harness-engineering, 2026-05-17

The local artifact confirms a mature docs surface rather than only a viral link: a dark lecture/library UI with top navigation for Lectures, Projects, Library, Skills, and Try Harness; lecture topics include "Why Capable Agents Still Fail," "What a Harness Actually Is," "Why Repository Must Become System of Record," "Why One Giant Instruction File Fails," and "Why Long-Running Tasks Lose Continuity." Source: local X artifact review, 2026-06-30

Companion: awesome-harness-engineering

The child repo walkinglabs/awesome-harness-engineering curates articles, playbooks, benchmarks, specs, and open-source projects for the practice. It is organized into Foundations; Context, Memory and Working State; Constraints, Guardrails and Safe Autonomy; Specs, Agent Files and Workflow Design; Evals and Observability; Benchmarks; and Runtimes, Harnesses and Reference Implementations. Generic agent tooling is explicitly out of scope unless it covers a reliability-critical harness primitive. Source: GitHub walkinglabs/awesome-harness-engineering, 2026-05-17

Current Source Snapshot

As of 2026-06-30, walkinglabs/learn-harness-engineering is MIT-licensed, default branch main, HEAD 2fdc46b20a3da696e1657a34fadfb5fd9e8acbb1, 9,436 stars, 1,003 forks, and has no GitHub releases or tags. walkinglabs/awesome-harness-engineering is default branch main, HEAD f84f1701974cf1ad67dd774b025b33e613275cee, 3,389 stars, 270 forks, and has no GitHub releases or tags. Brin returned 65 caution for both repositories, so use the course/list as learning and source discovery, not as executable dependency code. Source: GitHub API, git ls-remote, Brin API, 2026-06-30

Relevance to Kevin

This is the canonical external curriculum for Harness Engineering, the concept Kevin's wiki already documents (five defense layers, the verification gap, the Implementation Notes Pattern, Cloudflare's vuln harness). The minimal-harness templates map directly onto Kevin's own stack: AGENTS.md as schema, skills as invariants, state.json as the progress checkpoint, and doctors as evaluators. See Claude Code Harness and Agent Operating System.

---

Timeline

• 2026-06-30 | Deep-reviewed the bookmark artifact, course UI, course repo, awesome-list repo, and Brin score. Updated the page with current source snapshots and kept the decision as curriculum/reference only; no new skill is needed because the active harness skills and capsules own execution. Source: X bookmark artifact audit, 2026-06-30
• 2026-05-17 | Vaishnavi (@_vmlops) called it "the best site on the internet to learn harness engineering" (4,028 likes, 7,367 bookmarks). Resolved to the Walking Labs course plus the 2,893-star awesome-harness-engineering list. Source: X/@_vmlops, 2026-05-17