patterns_skill

Overview

This skill documents design patterns for building agents with the Langroid multi-agent LLM framework. It focuses on agent configuration, tool handlers, task control, batch workflows, and integrations with MCP servers and CLI coding agents. The patterns are practical, implementation-focused, and aimed at reliable, maintainable agent behavior.

How this skill works

The skill presents clear patterns you can apply when designing Langroid agents and tasks. Each pattern explains the objective, when to use it, and the concrete mechanism (agent methods, TaskConfig, run_batch_tasks, mcp_tool integration, quiet_mode). Examples show how to wire tools, validate outputs, manage state, and control task termination. Patterns are framework-agnostic but include actionable Python idioms for common scenarios.

When to use it

• When you need a simple agent that returns a structured tool response directly.
• When tool handlers must access external resources or maintain retryable state.
• When tool outputs require validation against input context before accepting.
• When you want to terminate a task only on a specific tool emission.
• When processing many similar inputs concurrently without manual concurrency management.
• When integrating external MCP tools (e.g., Claude Code) or silencing agent noise for clean CLI output.

Best practices

• Define stateful tool handlers as agent methods when you need shared resources, counters, or deterministic retries.
• Return plain error strings from handlers to trigger automatic LLM retries; use AgentDoneTool/DoneTool appropriately to signal success.
• Use TaskConfig.done_sequences to control exactly which tool emission ends a task.
• Use run_batch_tasks for isolation between items and to limit concurrency via batch_size to avoid connection exhaustion.
• Validate critical fields in handler methods and return clear error messages so the LLM can self-correct.
• Wrap long-running runs in quiet_mode() to suppress framework noise and print your own progress messages.

Example use cases

• A CLI code agent that edits and writes files via MCP tools exposed by Claude Code, with post-processing of tool results.
• A task that must preserve placeholders from the input; a handler validates and forces LLM retries until valid.
• Batch-linting or refactoring of hundreds of files using the same agent logic, executed in parallel with run_batch_tasks.
• A test harness that terminates only when a specific verification tool emits a pass signal via TaskConfig.done_sequences.
• A small agent that returns a single structured ToolMessage as the final output for downstream processing.