What is harness engineering, and how do I use it with AI coding agents?
In brief
Harness engineering designs the environment around an AI coding agent: context, tools, permissions and feedback that make progress verifiable. Start with one bounded task, durable project notes, an isolated workspace and checks tied to the delivered revision. Improve that path before adding orchestration.
- Make context survive a session reset.
- Provide an isolated, reproducible environment.
- Tie trustworthy checks to the delivered revision.
The problem: an agent needs more than a good prompt
Harness engineering designs the environment around an AI coding agent: context, tools, permissions and feedback. Plausible code can leave the replacement session unable to resume or the reviewer unable to verify it.
The challenge is making progress reproducible across sessions. Reliable feedback must test the requested behavior and remain trustworthy even when the agent can edit its own checks.
What you may have tried
- Better prompts: clarify instructions, but context can disappear at handoff.
- Larger context bundles: add information, but stale material can obscure the current task.
- More orchestration: routes work without proving correctness.
Start with a bounded task, durable notes, an isolated workspace and independent checks. A green build cannot prove a signup error preserves fields or creates no registration. An evaluator changed to always pass cannot establish correctness.
How Arcadia makes intent and feedback durable
Arcadia connects a Project’s Outcome to a Plan and current Action, with repository-held context, isolated sessions and revision-bound completion evidence. Arcadia currently drives Claude Code, Codex, and OpenCode. Coding agents remain interchangeable execution tools under your direction. This supplies structure around the coding agent rather than another code-generating loop.
For a signup repair, the Action names the invalid-email behavior. A fresh session can recover that task from project files and reproduce the check against the candidate revision.
The benefit is a recoverable task with inspectable proof. Test one interrupted task before adding more automation; engineering judgment and independent verification remain necessary.
Sources and further reading
- OpenAI: harness engineering — environments, intent and feedback around coding agents.
- Anthropic: effective harnesses for long-running agents — continuity across context windows.
- Anthropic: harness design for long-running applications — generation and evaluation as separate responsibilities.
- Arcadia’s preservation check repair — a concrete evaluator-binding case and its review evidence.
Current capabilities and supporting evidence
- Completion evidenceWorking
Acceptance criteria are checked before an action can be marked done.
Supporting evidence - Independent pull-request reviewWorking
Independent review is required before merge.
Supporting evidence - Agent session managementWorking
Isolated worktrees per action, across Codex, Claude Code and OpenCode.
Supporting evidence
Capability snapshot: . See what changed →