agent-frameworks-specialist

Agent Frameworks Specialist

When to Apply

• Building AI agents or multi-agent systems
• Integrating LLMs with tools, APIs, or external services
• Designing orchestration, routing, or handoff logic
• Implementing memory, state, or conversation persistence
• Adding observability to agent systems

Framework Quick Reference

Framework	Priority	Best For
Claude Agent SDK	CRITICAL	Claude-native agents, subagents, custom tools
OpenAI Agents SDK	CRITICAL	Multi-agent handoffs, guardrails, tool search
LangGraph	HIGH	Stateful graphs, conditional routing
Deep Agents (deepagents)	HIGH	Planning + subagents + virtual filesystem on top of LangGraph
Strands	HIGH	Model-agnostic, lightweight, AWS/Bedrock
CrewAI	MEDIUM	Role-based multi-agent teams
Mastra	MEDIUM	TypeScript-first, observational memory

For detailed code examples, read from rules/:

• rules/claude-sdk.md -- agentic loop, query(), subagents, custom tools
• rules/openai-sdk.md -- agents, handoffs, tool search, guardrails
• rules/langgraph.md -- StateGraph, Functional API, conditional routing
• rules/deepagents.md -- createDeepAgent, subagents, virtual filesystem, tool factory pattern
• rules/strands.md -- model providers, memory, MCP support
• rules/patterns.md -- architecture (ARCH-01-05), wiring, CrewAI, Mastra

Rules

• AF-R01 (CRITICAL): Use the agentic loop -- keep calling until stop signal.
• AF-R02 (CRITICAL): Tool descriptions are prompts. Write them clearly.
• AF-R03 (CRITICAL): No secrets in code. Use environment variables.
• AF-R04 (HIGH): Add observability from day one.
• AF-R05 (HIGH): Start single-agent. Add multi-agent only when needed.
• AF-R06 (HIGH): Use Context7 MCP for current SDK docs.
• AF-R07 (HIGH): Use tool search when tool count exceeds ~20.
• AF-R08 (MEDIUM): Test with temperature: 0 for reproducibility.
• AF-R09 (MEDIUM): Set token/cost budgets to prevent runaway loops.

Anti-Patterns

• God agent -- 50 tools in one agent. Split or use tool search.
• No stop condition -- always set max iterations.
• Synchronous everything -- use async/streaming.
• Logging messages only -- trace tool calls, latencies, tokens.
• Hardcoded model -- make model configurable.
• All tools upfront -- use defer_loading / ToolSearchTool.