Developing AI Agents plugin
A plugin with one skill for designing, debugging and evaluating AI agents, based on Bojie Li book on agent engineering.
- 51 ★ on the repository
- Last commit
- Apache-2.0
Minimal access
Why this level
- The skill holds instructions and the reference text of the book and runs nothing itself.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/developing-ai-agentsDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skill holds instructions and the reference text of the book and runs nothing itself.
This is third-party code. Review the repository files before installing.
What it does
The plugin installs one skill into Claude Code and Codex that helps design, fix and evaluate AI agents. It rests on the book by Bojie Li on agent design principles and engineering practice, second edition, and the Russian text of the book ships inside the plugin. The skill activates by itself when you describe a task in plain text. It helps decide whether an agent is needed or one model call or a plain scripted flow suffices, expose architecture weak spots and ways to test them, diagnose loops, lost context and repeated tool calls, and choose between one agent and several. To check whether a new prompt, model or tool improved results, the repository holds benchmarks from several iterations.
Who it is for. Engineers who design and debug agents on AI models and want systematic practice.
Good fit when
- You need to decide whether one model call is enough for a task such as invoice dispute handling.
- An agent forgets constraints and repeats calls after context compaction.
- You are choosing between one agent and a team of several.
- You want to check whether a new prompt or model improved results.
Not a fit when
- An ordinary application with no agent loop.
- You need a ready agent framework with code.
- You need the whole book in English.
Example request
I need an agent that handles invoice disputes. Which architecture is minimally sufficient and where should its permissions be limited?Limitations
The skill contains the Russian text of the book and follows its approach, so advice reflects the book author view. It only guides reasoning and checks and never runs your agents. It is unnecessary for ordinary applications with no agent loop. Start a new session after installing or updating. The benchmarks in the repository were run by the plugin author, and the README states no independent verification. The plugin includes the Russian text of the book and the README names no other language.
How to disable. Remove the plugin with `claude plugin uninstall` or the `codex plugin` command for your agent and end the session.
FAQ
How do I install the plugin in Claude Code?
Run `claude plugin marketplace add ilkruglov/developing-ai-agents-skill`, then `claude plugin install developing-ai-agents@developing-ai-agents-skill`.
How do I install it in Codex?
Run `codex plugin marketplace add ilkruglov/developing-ai-agents-skill`, then `codex plugin add developing-ai-agents@developing-ai-agents-skill`.
How do I invoke the skill?
Describe the task in plain text and the skill attaches itself. No separate command is needed.
How do I update the plugin?
For Claude Code run `claude plugin marketplace update developing-ai-agents-skill` and `claude plugin update developing-ai-agents@developing-ai-agents-skill`, then start a new session.
When is the skill unnecessary?
For an ordinary application with no agent loop; the README says so directly.
Discussion
No questions or reviews yet.
Related
PluginsSix plugins for Claude Code and Gemini CLI: 44 skills on designing AI products, from conversations to evaluation and prompts
AI Builder Club Skills
AI Builder Club — Skills
A Claude Code plugin marketplace to make a repo agent-ready and run compounding agent loops
Claude Code plugin with 60 subagents and 62 team presets, plus a browser designer for planning agent teams visually.
AI Agents Orchestrator
AI-Agents-Orchestrator
The AI Collaboration plugin hands a task between Claude Code, Codex, Cursor, Gemini CLI and Copilot via a shared SQLite history.
Agents Best Practices
agents-best-practices
A provider-neutral skill for designing agent systems: loop architecture, tools, permissions, memory and observability
A tool that builds tested skills from a real workflow and governs their lifecycle through your own marketplace
README in short
The plugin packages one skill and the Russian text of the book. The README lists situations where the skill helps and install commands for Claude Code and Codex.
README badge
Original SKILL.md
--- name: developing-ai-agents description: Use when designing, implementing, reviewing, debugging, or evaluating an AI agent or the environment that steers one - a skill, system prompt, CLAUDE.md or AGENTS.md, hook, tool or MCP server, subagent - especially for context engineering, tool interfaces, Harness reliability, memory, evaluation, post-training, continual evolution or self-evolution, asynchronous or realtime interaction, computer use, or multi-agent coordination. Do not use for ordinary non-agent application code. Not for AI infrastructure — GPU and serving capacity sizing, training clusters, or serving cost per token on given hardware (use designing-ai-infra); the agent's own logic, prompts, tools and memory, and its context, token use, cost and latency as design and eval concerns, stay here. --- # Разработка AI-агентов Проектируй агента как проверяемую систему, а не как «промпт плюс мощная модель». Опирайся на русское издание книги Bojie Li «AI-агенты изнутри: принципы проектирования и инженерная практика» и на факты из текущего проекта. ## Рабочий контракт 1. Сначала изучи реальные требования, код, конфигурацию, трассы, метрики и ограничения. Не называй гипотезу причиной без доказательств. 2. Разделяй: - **evidence** — что подтверждено кодом, логом, тестом или измерением; - **inference** — наиболее вероятное объяснение; - **unknown** — что еще нужно измерить. 3. Начинай ответ с решения или диагноза. После него дай минимальную архитектуру, порядок реализации, проверки и риски. 4. Для review или diagnosis оставайся read-only, пока пользователь явно не попросит изменить систему. Для реализации соблюдай инструкции репозитория и TDD. 5. Книжные принципы считай устойчивыми инженерными эвристиками. Актуальность моделей, SDK, API, протоколов, цен, лимитов и поддержки провайдеров проверяй по первичным текущим источникам. 6. По умолчанию укладывай архитектурный ответ в 1200 слов. Расширяй его только по явному запросу; если деталей больше, приоритизируй решения, contracts и проверки, а не полный checklist. ## С чего начать: маршрут по задаче | Запрос выглядит как | Начни с | Добери при необходимости | |---|---|---| | спроектировать агента с нуля | [playbooks/design-agent.md](references/playbooks/design-agent.md) | [templates/agent-design.md](references/templates/agent-design.md), [patterns.md](references/patterns.md) | | разобрать trace, найти причину сбоя | [playbooks/diagnose-trace.md](references/playbooks/diagnose-trace.md) | [templates/trace-diagnosis.md](references/templates/trace-diagnosis.md), [antipatterns.md](references/antipatterns.md) | | review существующей агентной системы | [playbooks/harness-review.md](references/playbooks/harness-review.md) | [templates/harness-spec.md](references/templates/harness-spec.md) | | построить или починить evals | [playbooks/build-evals.md](references/playbooks/build-evals.md) | [templates/eval-plan.md](references/templates/eval-plan.md) | | память, RAG, самоулучшение | [playbooks/memory-design.md](references/playbooks/memory-design.md) | [templates/memory-policy.md](references/templates/memory-policy.md) | | голос, realtime, асинхронность, Computer Use | [playbooks/realtime-latency.md](references/playbooks/realtime-latency.md) | [chapters/ch06](references/chapters/ch06-interaction.md) | | один агент или несколько | [playbooks/multi-agent-choice.md](references/playbooks/multi-agent-choice.md) | [chapters/ch10](references/chapters/ch10-multi-agent.md) | | контракт инструмента, права, песочница | [templates/tool-contract.md](references/templates/tool-contract.md) | [chapters/ch04](references/chapters/ch04-tools.md) | | симптом известен, причина нет | [antipatterns.md](references/antipatterns.md) | playbook по нужной области | | что говорит книга по теме | [source-map.md](references/source-map.md) | нужный конспект главы | | спорный вопрос проектирования, похожий на «вопрос для размышления» | [chapters/ch12](references/chapters/ch12-reference-answers.md) | ответ в `references/source-book/reference-answers.md` | Загружай только релевантные файлы. Не помещай всю книгу в контекст одновременно. ## Основная модель > **AI-агент = LLM + контекст + инструменты** - **LLM** принимает решения, но не хранит надежное состояние системы. - **Контекст** определяет, что агент видит сейчас: инструкции, состояние, историю и результаты действий; определения инструментов входят сюда же и занимают тот же бюджет.