Autonomous Agent Harness
Designs autonomous agent loops that remain controlled, recoverable, and auditable.
The skill explicitly separates locked, editable, append-only, and human-controlled surfaces, and requires rollback, durable logs, and human approval for merge and deployment. It does not address sensitive-data handling, credential isolation, dependency security, general data-flow disclosure, or recovery procedures in sufficient detail, so points are deducted.
The instructions are internally coherent and cover keep, discard, crash, timeout, and review states with useful loops and checklists. However, the skill contains no executable scripts, tests, or concrete abnormal-input and failure-message examples, so key paths cannot be reproduced statically; the score remains conservative under the static cap.
Activation scenarios are specific, and adjacent non-fit areas are explicitly assigned to other skills such as evaluation, tool-design, project-development, and hosted-agents. It does not define standard inputs and outputs, broader non-fit boundaries, Chinese-language usage, or mainland-China environment compatibility, so points are deducted.
The document is well organized, with concepts, detailed topics, practical checklists, file layout, examples, gotchas, and integrations. Repository evidence supplies an MIT license, while the skill includes creation date, update date, author, and version. It lacks a skill-specific changelog, clearly assigned maintenance responsibility, and an explicit update path; referenced internal documents were not supplied in the evidence.
Guidance on locked evaluators, narrow edit surfaces, durable logs, novelty gates, and human approval is directly actionable for harness design, and the examples cover representative patterns. The output is conceptual rather than runnable, and cited internal runbooks are unavailable here, leaving substantial implementation and validation work; the static cap therefore applies.
The skill names several external research and engineering references, providing limited traceability. It supplies no precise citations or locations and no committed tests or evidence covering this skill’s key paths; nothing was executed in this review, so only limited credit is justified.
- This is a low-confidence static review; no scripts, tests, or examples were executed.
- Implementation should add explicit credential and sensitive-data isolation, network and filesystem permissions, crash and timeout recovery, rollback boundaries, and user-confirmation rules.
- External resources and internal references were not individually verified in the supplied material; repository-level CI claims should not be treated as executed validation of this skill.
What does this skill do, and when should you use it?
Harness Engineering is for designing the control system around an agent: its editable surfaces, feedback channels, durable state, recovery behavior, and approval boundaries. It covers research loops, evaluation scaffolds, locked and editable surfaces, durable logs, novelty gates, pruning, rollback, PR preparation, and human review. The skill is most relevant when agents must operate for long periods without weakening their own evaluators or taking irreversible actions. The repository is MIT-licensed and distributes this skill alongside 22 other separate skills.
Helps divide an agent environment into locked, editable, append-only, and human-controlled surfaces; define keep, discard, crash, timeout, rollback, and review states; persist plans, source queues, results, failures, and handoffs in files; add source retrieval, novelty checks, mechanism registries, and pruning to research-to-skill workflows; and design monitoring reports plus PR preparation that stops short of automatic merge or deployment.
- A research team is building an autonomous loop that retrieves sources, evaluates candidates, and proposes skill changes.
- An engineering team wants an agent to optimize code or prompts while keeping evaluation scripts and metrics locked.
- A platform team needs background agents to run for hours or days and recover from compaction or interruptions through thread and result logs.
- A maintainer wants an agent to prepare a branch, run checks, and draft a PR while retaining merge, deployment, and destructive actions for human approval.
- A system designer needs to detect duplicate mechanisms, preserve failed attempts, and prune growing experiment stacks.
What are this skill's strengths and limitations?
- Provides a clear model for separating agent-editable content from locked and human-controlled surfaces.
- Covers durable logs, rejected attempts, rollback, novelty checks, and pruning for long-running loops.
- Includes concrete patterns for autoresearch-style optimization, research-to-skill workflows, and auto-PR preparation.
- Addresses metric gaming, missing retrieval evidence, stale monitoring, and ambiguous approval boundaries.
- It primarily provides design guidance, workflows, and pseudocode; the skill content does not include a directly runnable implementation.
- It does not provide complete configuration examples for a specific client, deployment environment, or runtime.
- No independent test suite or measured performance evidence is shown for this individual skill, so target-environment validation remains necessary.
- It discusses source refresh and file-based state but does not prescribe a particular network service, storage backend, or permission model.
How do you install this skill?
Add the repository as a plugin directory, or copy the skill directory into a project skill directory. For a Codex project, create .codex/skills/ and copy skills/harness-engineering into it; Claude Code projects can use .claude/skills/, and Cursor projects can use .cursor/skills/. Keep the directory layout intact rather than flattening SKILL.md into one file. The repository is licensed under MIT.
How do you use this skill?
In an Agent Skills-compatible client, request a harness-design task such as: “Design an autonomous research loop with a locked evaluator, append-only results log, failure rollback, and human merge approval.” The skill then frames the design around editable surfaces, feedback, persistence, and governance. A client-specific activation command is not documented in the skill.
How does this skill compare with similar options?
This skill differs from the repository’s evaluation skill, which covers general evaluation frameworks and quality gates; Harness Engineering focuses on autonomous control surfaces, persistence, rollback, and approval. It is also distinct from tool-design, which covers tool contracts and recovery errors, and hosted-agents, which covers remote sandbox and hosted-session infrastructure.