Historical source material recovered on 2026-09-09. Capture dates and source limitations apply; recovery is not fresh verification.

Harness Engineering

Harness engineering is the practice of making a codebase, toolchain, and team process legible enough that coding agents can work, verify, review, and hand off changes with less human babysitting.

Key points

• OpenAI frames the move from autocomplete and pair programming into agent delegation as requiring a better "harness" around the model, not only a better model ^[src-084].
• The API & Codex Build Hour demo builds an "agent legibility score" for repositories, with criteria such as bootstrap self-sufficiency, task entry points, validation harnesses, linting, formatting, documentation, and modular boundaries ^[src-084].
• Worktrees are part of the harness because they let multiple agent tasks run in parallel on separate branches without clobbering each other ^[src-084].
• High-quality validation is central: tests, lint, build commands, browser checks, logs, and other acceptance signals let Codex decide whether the work is actually done ^[src-084].
• Skills and subagents encode repeatable team standards, such as PR creation, commit style, code review, architecture review, standards enforcement, and pull-request follow-up ^[src-084].
• Team knowledge should move from individual heads and chat history into version-controlled files, docs, specs, skills, notes, and local decision records so every agent benefits from the same context ^[src-084].
• The source also adds a company-context version: an "Atlas" style repo for operating principles, strategy, and non-code context can make Codex useful beyond production code ^[src-084].
• ^[src-094] defines the harness as the surrounding machinery that turns a raw model into a working agent: prompts, tools, context policies, hooks, sandboxes, sub-agents, orchestration, observability, and constraints.
• The paper warns against treating the model as the whole system; agent behaviour is heavily shaped by the harness around the model, not only by model capability ^[src-094].
• In the AI-assisted SDLC, the harness is configured during requirements, planning, and architecture; run during implementation; improved through testing and QA; and observed during review, deployment, and maintenance ^[src-094].
• GitHub's worktree guide gives the low-level filesystem pattern for safe parallel coding work: separate working directories let humans and agents work on different branches without stashing or disturbing editor state ^[src-098].
• Google's ADK guide gives the long-running workflow pattern for production agents: persistent sessions, state machines, event-driven dormancy, and multi-agent delegation are part of the harness, not optional polish ^[src-101].
• ^[src-202] adds fresh World's Fair evidence for harness-over-model thinking: liveness models, save buttons, multi-machine fleets, physical-data harnesses, process discipline, and adaptive engineering all become part of what makes an agent reliable.
• The playlist also connects harness engineering to UX and review: a good harness should produce receipts, make "done" inspectable, and reduce the hidden review debt created by coding agents ^[src-202].

Related entities

• Openai
• Codex
• Gpt 54

Related concepts

• Agentic Engineering
• Agent Harness Portability
• Codex Project Workspace
• Context Engineering
• Progressive Context Loading
• Human Agent Collaboration
• Continuous Agent Evaluation
• Vibe Coding
• Software Factory Model
• Token Economics
• Git Worktrees For Parallel Agents
• Long Running Agents
• Google Agent Development Kit
• Agentic CI/CD
• Agent Evidence Receipts

Source references

• ^[src-084] OpenAI Codex, Workspace Agents, Prompt Caching, and Superintelligence Policy cluster (2026-02-09 to 2026-05-08)
• ^[src-094] Addy Osmani, Shubham Saboo, Sokratis Kartakis – "The New SDLC With Vibe Coding" (2026-05)
• ^[src-098] Cassidy Williams – "What are git worktrees, and why should I use them?" (2026-06-16)
• ^[src-101] Shubham Saboo and Eric Dong – "Build Long-running AI agents that pause, resume, and never lose context with ADK" (2026-05-12)
• ^[src-202] AI Engineer World's Fair Online Track 2026 playlist update (47 new transcript captures, 2026-06-22 to 2026-07-02)

Historical recovered snapshot — 2026-09-07

Preserved from manual-three-20260907. Documentary recovery on 2026-09-09; claims below retain their historical source scope and were not freshly verified. Current synthesis above remains authoritative.

GitHub distinguishes a harness (tools, permissions, memory, context and orchestration), loops (repeatable workflows), squads (specialised roles), fleets (parallel agents), and hill climbing (evaluation-driven improvement). These definitions guide system design but establish no measured reliability gain. ^[src-404]

• GitHub Podcast
• Article summary

Scope: this pass verifies the added source; earlier cited evidence is retained, not independently re-audited.

Evidence checked on 14 September 2026

Google describes a reusable Antigravity runtime with persistence, declarative policies and lifecycle hooks. The example combines an agent core with telemetry middleware for a multi-agent control plane. ^[src-450] Architecture guidance and examples; production reliability was not independently measured.

• Power agent hubs or custom harnesses with the Antigravity SDK

Google recommends behavioural evaluations that inspect observable tool calls and intermediate actions to diagnose regressions beyond end-to-end task scores. ^[src-467] Design guidance rather than proof of improvement; deterministic assertions check traces, not deterministic model behaviour.

• The Anatomy of Harness Engineering