agent-harness-design
Designing agent harnesses and tool systems — risk taxonomy for tools, permission decisions, draft/commit pattern, structured tool results, agent budgets (10 types), context trust labels against prompt injection, plan-artifact, approval records, observability and traces, evals (13 categories), event model, streaming buffering, 3rd-party skill install checklist, agentic RAG, self-improving SOP loops, model policy, reasoning effort, and Programmatic Tool Calling adoption gates. Use when building a new Agent SDK app, custom orchestrator, MCP server, Cloudflare Worker with tool calls, agentic RAG pipeline, model router, or model-tier policy; when designing tools and permissions; when writing an agent loop; or when you need trust labels for external content. Do NOT use for improving or auditing an already-built harness (use harness-audit / harness-design instead), nor for ordinary Claude Code sessions where the harness is already given.
- License 许可
- MIT
- 最近更新
- 2026-11-11
- Stars
- ★ 214
- Forks
- 32
01 前言 Front Matter
- name
- agent-harness-design
- description
- Designing agent harnesses and tool systems — risk taxonomy for tools, permission decisions, draft/commit pattern, structured tool results, agent budgets (10 types), context trust labels against prompt injection, plan-artifact, approval records, observability and traces, evals (13 categories), event model, streaming buffering, 3rd-party skill install checklist, agentic RAG, self-improving SOP loops, model policy, reasoning effort, and Programmatic Tool Calling adoption gates. Use when building a new Agent SDK app, custom orchestrator, MCP server, Cloudflare Worker with tool calls, agentic RAG pipeline, model router, or model-tier policy; when designing tools and permissions; when writing an agent loop; or when you need trust labels for external content. Do NOT use for improving or auditing an already-built harness (use harness-audit / harness-design instead), nor for ordinary Claude Code sessions where the harness is already given.
02 资源文件 11
- 参考文档references/agent-approval-records.md6.4 KB
- 参考文档references/agent-budgets.md6.3 KB
- 参考文档references/agent-evals.md7.5 KB
- 参考文档references/agent-event-model.md7.3 KB
- 参考文档references/agent-observability.md6.5 KB
- 参考文档references/agent-plan-artifact.md7.2 KB
- 参考文档references/agent-skill-install-checklist.md14.9 KB
- 参考文档references/agent-streaming.md6.7 KB
- 参考文档references/agent-tool-design.md17.4 KB
- 参考文档references/agentic-rag-model-policy.md7.8 KB
- 参考文档references/context-trust-labels.md9.2 KB
03 SKILL.md 正文
Agent Harness Design
Eleven operational reference sheets for designing a safe, observable agent harness. They are situational — load only the one(s) relevant to the current task from references/ (this is why they live in a skill rather than always-on rules: building an agent harness is occasional, so the detail should not bloat every session's context).
references/agent-tool-design.md— 15-class risk taxonomy, 7-type permission decision object, draft/commit naming, structured tool results, deferred tool loading, hosted vs client tools, connector code-execution pattern.references/context-trust-labels.md— trusted / semi_trusted / untrusted labels + verbatim boundary statement; prompt-injection defense.references/agent-budgets.md— 10 mandatory budget types every agent loop must declare.references/agent-evals.md— 13 eval categories + 13 adversarial test cases + when to add regression evals.references/agent-observability.md— 16 trace fields per model call, 7-question audit, 6-step incident response.references/agentic-rag-model-policy.md— self-improving agentic RAG state, specialist roles, evaluation vectors, Pareto selection, OpenAI model/effort policy, and Programmatic Tool Calling adoption gates.references/agent-plan-artifact.md— planning mode, plan artifact format (10 fields), plan-validate-execute.references/agent-approval-records.md— approval request/result JSON schemas, scope/expiration, no self-approval.references/agent-streaming.md— buffering for incremental tool calls when stream=True; abort handling; output guardrail modes.references/agent-event-model.md— 13 typed events for harness state persistence (replay/audit/compaction/evals).references/agent-skill-install-checklist.md— pre/during/post install + audit + incident response for 3rd-party skills.
Source: distilled from the agents-best-practices skill (Denis Sergeevitch, MIT) + Anthropic harness-design engineering. Read the specific reference before applying — do not work from this index alone.
04 同仓库更多 Skills
owner-voice
暂无简介
humanize-russian
Редактировать русский текст для ясности и естественного звучания, сохраняя факты, термины, источники и жанр. Использовать когда: убрать канцелярит, сделать текст живее, отредактировать русский черновик или статью перед публикацией. Не определяет авторство и не обещает обход детекторов. Для английского — humanize-english; для структуры и доказательств — article-structure-review.
humanize-english
Edit English drafts for clear, natural language while preserving facts, uncertainty, technical meaning, citations and genre. Use when: humanize this text, remove AI-style filler, improve English prose, or edit an English article before publication. Not an authorship detector or a promise to bypass one. For Russian use humanize-russian; for thesis/evidence and structure use article-structure-review.
caveman
Ultra-compressed reply mode. Cuts output tokens ~65-75% by answering in terse "caveman" style while keeping full technical accuracy, and shapes what survives: action first, state restated every turn, concrete time estimates, visible wins. Intensity levels: lite, full (default), ultra, wenyan-lite, wenyan-full, wenyan-ultra. Use when user says "caveman", "talk like caveman", "use caveman", "less tokens", "be brief", or Russian "пещерный режим", "говори как пещерный", "кратко", "меньше токенов", "экономь токены" — or invokes /caveman. Preserves the user's language (Russian in → Russian caveman out). Off only: "stop caveman" / "normal mode" / "обычный режим". Do not use merely because a response should be concise, and do not use when the user requests normal prose, a polished document, or exact wording.
article-structure-review
Проверить структуру технической статьи после черновика: задача читателя, жанр, опора значимых утверждений на доказательства и честные ограничения. Use after the first draft, before publication. Исправлять структуру, когда это входит в запрос; не создавать обязательный конвейер из других навыков. Для локальной языковой редактуры — humanize-english или humanize-russian.
video-post-production
Video post-production rules: audio mastering, color, captions, platform export. Use when: 'add music', 'add voiceover', 'export for tiktok', 'add captions', 'color grade', 'audio levels', 'master audio', 'export settings', 'platform requirements'. Covers FFmpeg patterns, audio chain, subtitle standards, and platform-specific export configs. Do NOT use to generate or animate the video frames/scenes in code (use remotion-production-guide), to write the script (use video-narrative-arc), or to evaluate script quality (use script-evaluator); this is the finishing stage applied to an already-rendered video.