ai-agents（AI 特工）

AI Agents Development — Production Skill Hub

Modern Best Practices (March 2026): deterministic control flow, bounded tools, auditable state, MCP-based tool integration, handoff-first orchestration, multi-layer guardrails, OpenTelemetry tracing, and human-in-the-loop controls (OWASP LLM Top 10: https://owasp.org/www-project-top-10-for-large-language-model-applications/).

This skill provides production-ready operational patterns for designing, building, evaluating, and deploying AI agents. It centralizes procedures, checklists, decision rules, and templates used across RAG agents, tool-using agents, OS agents, and multi-agent systems.

No theory. No narrative. Only operational steps and templates.

When to Use This Skill

Codex should activate this skill whenever the user asks for:

• Designing an agent (LLM-based, tool-based, OS-based, or multi-agent).
• Scoping capability maturity and rollout risk for new agent behaviors.
• Creating action loops, plans, workflows, or delegation logic.
• Writing tool definitions, MCP tools, schemas, or validation logic.
• Generating RAG pipelines, retrieval modules, or context injection.
• Building memory systems (session, long-term, episodic, task).
• Creating evaluation harnesses, observability plans, or safety gates.
• Preparing CI/CD, rollout, deployment, or production operational specs.
• Producing any template in /references/ or /assets/.
• Implementing MCP servers or integrating Model Context Protocol.
• Setting up agent handoffs and orchestration patterns.
• Configuring multi-layer guardrails and safety controls.
• Evaluating whether to build an agent (build vs not decision).
• Calculating agent ROI, token costs, or cost/benefit analysis.
• Assessing hallucination risk and mitigation strategies.
• Deciding when to kill an agent project (kill triggers).
• For prompt scaffolds, retrieval tuning, or security depth, see Scope Boundaries below.

Scope Boundaries (Use These Skills for Depth)

• Prompt scaffolds & structured outputs → ai-prompt-engineering
• RAG retrieval & chunking → ai-rag
• Search tuning (BM25/HNSW/hybrid) → ai-rag
• Security/guardrails → ai-mlops
• Inference optimization → ai-llm-inference

Default Workflow (Production)

• Pick an architecture with the Decision Tree (below); default to workflow/FSM/DAG for production.
• Draft an agent spec with assets/core/agent-template-standard.md (or assets/core/agent-template-quick.md).
• Specify tools and handoffs with JSON Schema using assets/tools/tool-definition.md and references/api-contracts-for-agents.md.
• Add retrieval only when needed; start with assets/rag/rag-basic.md and scale via assets/rag/rag-advanced.md + references/rag-patterns.md.
• Add eval + telemetry early via references/evaluation-and-observability.md.
• Run the go/no-go gate with assets/checklists/agent-safety-checklist.md.
• Plan deploy/rollback and safety controls via references/deployment-ci-cd-and-safety.md.

Quick Reference

Framework Selection (March 2026)

Tier 1 — Production-Grade

Tier 2 — Specialized

Tier 3 — Niche

See references/modern-best-practices.md for detailed comparison and selection guide.

Framework Deep Dives

• Claude Agent SDK - references/claude-agent-sdk-patterns.md Agent definition, built-in tools (Bash, TextEditor, Computer), MCP servers, guardrails, multi-agent, streaming events
• Pydantic AI - references/pydantic-ai-patterns.md Type-safe agents, MCP toolsets, native A2A, pydantic-graph FSM, durable execution, HITL, TestModel testing

Decision Tree: Choosing Agent Architecture

What does the agent need to do?
    ├─ Answer questions from knowledge base?
    │   ├─ Simple lookup? → RAG Agent (LangChain/LlamaIndex + vector DB)
    │   └─ Complex multi-step? → Agentic RAG (iterative retrieval + reasoning)
    │
    ├─ Perform external actions (APIs, tools, functions)?
    │   ├─ 1-3 tools, linear flow? → Tool-Using Agent (LangGraph + MCP)
    │   └─ Complex workflows, branching? → Planning Agent (ReAct/Plan-Execute)
    │
    ├─ Write/modify code autonomously?
    │   ├─ Single file edits? → Tool-Using Agent with code tools
    │   └─ Multi-file, issue resolution? → Code/SWE Agent (HyperAgent pattern)
    │
    ├─ Delegate tasks to specialists?
    │   ├─ Fixed workflow? → Multi-Agent Sequential (A → B → C)
    │   ├─ Manager-Worker? → Multi-Agent Hierarchical (Manager + Workers)
    │   └─ Dynamic routing? → Multi-Agent Group Chat (collaborative)
    │
    ├─ Control desktop/browser?
    │   └─ OS Agent (Anthropic Computer Use + MCP for system access)
    │
    └─ Hybrid (combination of above)?
        └─ Planning Agent that coordinates:
            - Tool-using for actions (MCP)
            - RAG for knowledge (MCP)
            - Multi-agent for delegation (A2A)
            - Code agents for implementation

Protocol Selection:

• Use MCP for: Tool access, data retrieval, single-agent integration
• Use A2A for: Agent-to-agent handoffs, multi-agent coordination, task delegation

Framework Selection (after choosing architecture):

Which framework?
    ├─ MVP/Prototyping?
    │   ├─ Python → OpenAI Agents SDK or CrewAI
    │   └─ TypeScript → Mastra or Claude Agent SDK
    │
    ├─ Production →
    │   ├─ Auditability/compliance? → LangGraph
    │   ├─ Type safety + MCP/A2A native? → Pydantic AI
    │   ├─ Anthropic models + Computer Use? → Claude Agent SDK
    │   ├─ Google Cloud / Gemini? → Google ADK
    │   ├─ Azure / .NET / Java? → MS Agent Framework
    │   ├─ AWS managed? → Bedrock Agents
    │   └─ RAG-heavy? → LlamaIndex Workflows
    │
    ├─ Minimalist / Research →
    │   ├─ Fewest LLM calls? → SmolAgents
    │   └─ Optimize prompts automatically? → DSPy
    │
    └─ Enterprise pipeline → Haystack

Core Concepts (Vendor-Agnostic)

Control Flow Options

• Reactive: direct tool routing per user request (fast, brittle if unbounded).
• Workflow (FSM/DAG): explicit states and transitions (default for deterministic production).
• Planner/Executor: plan with strict budgets, then execute step-by-step (use when branching is unavoidable).
• Orchestrated multi-agent: separate roles with validated handoffs (use when specialization is required).

Memory Types (Tradeoffs)

• Short-term (session): cheap, ephemeral; best for conversational continuity.
• Episodic (task): scoped to a case/ticket; supports audit and replay.
• Long-term (profile/knowledge): high risk; requires consent, retention limits, and provenance.

Failure Handling (Production Defaults)

• Classify errors: retriable vs fatal vs needs-human.
• Bound retries: max attempts, backoff, jitter; avoid retry storms.
• Fallbacks: degraded mode, smaller model, cached answers, or safe refusal.

Do / Avoid

Do

• Do keep state explicit and serializable (replayable runs).
• Do enforce tool allowlists, scopes, and idempotency for side effects.
• Do log traces/metrics for model calls and tool calls (OpenTelemetry GenAI semantic conventions: https://opentelemetry.io/docs/specs/semconv/gen-ai/).

Avoid

• Avoid runaway autonomy (unbounded loops or step counts).
• Avoid hidden state (implicit memory that cannot be audited).
• Avoid untrusted tool outputs without validation/sanitization.

Navigation: Economics & Decision Framework

Should You Build an Agent?

• Build vs Not Decision Framework - references/build-vs-not-decision.md

- 10-second test (volume, cost, error tolerance) - Red flags and immediate disqualifiers - Alternatives to agents (usually better) - Full decision tree with stage gates - Kill triggers during development and post-launch - Pre-build validation checklist

Agent ROI & Token Economics

• Agent Economics - references/agent-economics.md

- Token pricing by model (January 2026) - Cost per task by agent type - ROI calculation formula and tiers - Hallucination cost framework and mitigation ROI - Investment decision matrix - Monthly tracking dashboard

Navigation: AI Engine Layers

Five-layer architecture for production agent systems. Start with the overview, then drill into layer-specific patterns.

• AI Engine Architecture — references/ai-engine-layers.md 5-layer composition model, layer interaction matrix, implementation phases
• Context Graph Patterns — references/context-graph-patterns.md Node/edge schema, traversal patterns, graph-RAG, memory tiers, conflict detection
• Inbox Engine Patterns — references/inbox-engine-patterns.md Event-driven intake, signal classification, deduplication, priority routing, dead letter
• Knowledge Base Architecture — assets/knowledge-base/kb-architecture.md Unified KB schema (vector + graph + doc index), provenance, freshness, multi-tenant

Action Graph → covered by references/operational-patterns.md + references/agent-operations-best-practices.md Data Agent → covered by ../ai-rag/SKILL.md + references/rag-patterns.md

Navigation: Core Concepts & Patterns

Governance & Maturity

• Agent Maturity & Governance - references/agent-maturity-governance.md

- Capability maturity levels (L0-L4) - Identity & policy enforcement - Fleet control and registry management - Deprecation rules and kill switches

Modern Best Practices

• Modern Best Practices - references/modern-best-practices.md

- Model Context Protocol (MCP) - Agent-to-Agent Protocol (A2A) - Agentic RAG (Dynamic Retrieval) - Multi-layer guardrails - LangGraph over LangChain - OpenTelemetry for agents

Context Management

• Context Engineering - references/context-engineering.md

- Progressive disclosure - Session management - Memory provenance - Retrieval timing - Multimodal context

Core Operational Patterns

• Operational Patterns - references/operational-patterns.md

- Agent loop pattern (PLAN → ACT → OBSERVE → UPDATE) - OS agent action loop - RAG pipeline pattern - Tool specification - Memory system pattern - Multi-agent workflow - Safety & guardrails - Observability - Evaluation patterns - Deployment & CI/CD

Navigation: Protocol Implementation

• MCP Practical Guide - references/mcp-practical-guide.md Building MCP servers, tool integration, and standardized data access
• MCP Server Builder - references/mcp-server-builder.md End-to-end checklist for workflow-focused MCP servers (design → build → test)
• A2A Handoff Patterns - references/a2a-handoff-patterns.md Agent-to-agent communication, task delegation, and coordination protocols
• Protocol Decision Tree - references/protocol-decision-tree.md When to use MCP vs A2A, decision framework, and selection criteria

Navigation: Agent Capabilities

• Agent Operations - references/agent-operations-best-practices.md Action loops, planning, observation, and execution patterns
• RAG Patterns - references/rag-patterns.md Contextual retrieval, agentic RAG, and hybrid search strategies
• Memory Systems - references/memory-systems.md Session, long-term, episodic, and task memory architectures
• Tool Design & Validation - references/tool-design-specs.md Tool schemas, validation, error handling, and MCP integration

Skill Packaging & Sharing

• Skill Lifecycle - references/skill-lifecycle.md Scaffold, validate, package, and share skills with teams (Slack-ready)
• API Contracts for Agents - references/api-contracts-for-agents.md Request/response envelopes, safety gates, streaming/async patterns, error taxonomy
• Multi-Agent Patterns - references/multi-agent-patterns.md Manager-worker, sequential, handoff, and group chat orchestration
• OS Agent Capabilities - references/os-agent-capabilities.md Desktop automation, UI grounding, and computer use patterns
• Code/SWE Agents - references/code-swe-agents.md SE 3.0 paradigm, autonomous coding patterns, SWE-Bench, HyperAgent architecture

Framework-Specific Patterns

• Pydantic AI Patterns - references/pydantic-ai-patterns.md Type-safe agents, MCP toolsets (Stdio/SSE/StreamableHTTP), A2A via to_a2a(), pydantic-graph FSM, durable execution, TestModel testing

Navigation: Production Operations

• Evaluation & Observability - references/evaluation-and-observability.md OpenTelemetry GenAI, metrics, LLM-as-judge, and monitoring
• Deployment, CI/CD & Safety - references/deployment-ci-cd-and-safety.md Multi-layer guardrails, HITL controls, NIST AI RMF, production checklists
• Agent Debugging Patterns - references/agent-debugging-patterns.md Systematic debugging for agentic systems: trace analysis, tool call failures, loop detection, state corruption
• Voice & Multimodal Agents - references/voice-multimodal-agents.md Voice-first and multimodal agent patterns: speech pipelines, vision grounding, cross-modal orchestration
• Guardrails Implementation - references/guardrails-implementation.md Multi-layer guardrail patterns: input/output validation, content filtering, PII detection, cost caps

Navigation: Templates (Copy-Paste Ready)

Checklists

• Agent Design & Safety Checklist - assets/checklists/agent-safety-checklist.md Go/No-Go safety gate: permissions, HITL triggers, eval gates, observability, rollback

Core Agent Templates

• Standard Agent Template - assets/core/agent-template-standard.md Full production spec: memory, tools, RAG, evaluation, observability, safety
• Specialized Agent Template - assets/core/agent-template-specialized.md Domain-specific agents with custom capabilities and constraints
• Quick Agent Template - assets/core/agent-template-quick.md Minimal viable agent for rapid prototyping

RAG Templates

• Basic RAG - assets/rag/rag-basic.md Simple retrieval-augmented generation pipeline
• Advanced RAG - assets/rag/rag-advanced.md Contextual retrieval, reranking, and agentic RAG patterns
• Hybrid Retrieval - assets/rag/hybrid-retrieval.md Semantic + keyword search with BM25 fusion

Tool Templates

• Tool Definition - assets/tools/tool-definition.md MCP-compatible tool schemas with validation and error handling
• Tool Validation Checklist - assets/tools/tool-validation-checklist.md Testing, security, and production readiness checks

Multi-Agent Templates

• Manager-Worker Template - assets/multi-agent/manager-worker-template.md Orchestration pattern with task delegation and result aggregation
• Evaluator-Router Template - assets/multi-agent/evaluator-router-template.md Dynamic routing with quality assessment and domain classification

Service Layer Templates

• FastAPI Agent Service - ../dev-api-design/assets/fastapi/fastapi-complete-api.md Auth, pagination, validation, error handling; extend with model lifespan loads, SSE, background tasks

External Sources Metadata

• Curated References - data/sources.json Authoritative sources spanning standards, protocols, and production agent frameworks

Shared Utilities (Centralized patterns — extract, don't duplicate)

• ../software-clean-code-standard/utilities/llm-utilities.md — Token counting, streaming, cost estimation
• ../software-clean-code-standard/utilities/error-handling.md — Effect Result types, correlation IDs
• ../software-clean-code-standard/utilities/resilience-utilities.md — p-retry v6, circuit breaker for API calls
• ../software-clean-code-standard/utilities/logging-utilities.md — pino v9 + OpenTelemetry integration
• ../software-clean-code-standard/utilities/observability-utilities.md — OpenTelemetry SDK, tracing, metrics
• ../software-clean-code-standard/utilities/testing-utilities.md — Test factories, fixtures, mocks
• ../software-clean-code-standard/references/clean-code-standard.md — Canonical clean code rules (CC-*) for citation

Trend Awareness Protocol

IMPORTANT: When users ask framework recommendations or "what's best for X" questions, use WebSearch to verify current landscape before answering. If unavailable, use data/sources.json and state what was verified vs assumed.

Trigger: framework comparisons, "best for [use case]", "is X still relevant?", "latest in AI agents", MCP server availability.

Report: current landscape, emerging trends, deprecated patterns, recommendation with rationale.

Related Skills

This skill integrates with complementary skills:

Core Dependencies

• ../ai-llm/ - LLM patterns, prompt engineering, and model selection for agents
• ../ai-rag/ - Deep RAG implementation: chunking, embedding, reranking
• ../ai-prompt-engineering/ - System prompt design, few-shot patterns, reasoning strategies

Production & Operations

• ../qa-observability/ - OpenTelemetry, metrics, distributed tracing
• ../software-security-appsec/ - OWASP Top 10, input validation, secure tool design
• ../ops-devops-platform/ - CI/CD pipelines, deployment strategies, infrastructure

Supporting Patterns

• ../dev-api-design/ - REST/GraphQL design for agent APIs and tool interfaces
• ../ai-mlops/ - Model deployment, monitoring, drift detection
• ../qa-debugging/ - Agent debugging, error analysis, root cause investigation
• ../dev-ai-coding-metrics/ - Team-level AI coding metrics: adoption, DORA/SPACE, ROI, DX surveys (this skill covers per-task agent economics)

Usage pattern: Start here for agent architecture, then reference specialized skills for deep implementation details.

Usage Notes

• Modern Standards: Default to MCP for tools, agentic RAG for retrieval, handoff-first for multi-agent
• Lightweight SKILL.md: Use this file for quick reference and navigation
• Drill-down resources: Reference detailed resources for implementation guidance
• Copy-paste templates: Use templates when the user asks for structured artifacts
• External sources: Reference data/sources.json for authoritative documentation links
• No theory: Never include theoretical explanations; only operational steps

AI-Native SDLC Template

• Use assets/agent-template-ainative-sdlc.md for the Delegate → Review → Own runbook (guardrails + outputs checklist).

Fact-Checking

• Use web search/web fetch to verify current external facts, versions, pricing, deadlines, regulations, or platform behavior before final answers.
• Prefer primary sources; report source links and dates for volatile information.
• If web access is unavailable, state the limitation and mark guidance as unverified.