Step-by-step, code-heavy guides for building AI agents.
37 posts
A runnable Python guide to Claude structured output: the native output_config.format JSON path, Pydantic-driven schemas, strict tool use, a portable tool-use fallback, and the 400/truncation gotchas nobody covers.
A 429 is your rate limit; a 529 is Anthropic overloaded. Catch both, honor retry-after, back off with jitter, cap the retries, and never retry a 400. Full runnable Python wrapper for agent loops.
GraphRecursionError means you hit LangGraph's 25-superstep recursion limit. Raise it the right way, find the real infinite loop, and exit gracefully with RemainingSteps.
The evaluator-optimizer workflow puts a rubric-driven critic in the loop so a generator revises until it passes. A runnable Python build with real guardrails.
Run independent LLM calls at the same time in Python with asyncio.gather, cap them with a Semaphore, and survive 429 rate limits with backoff. Runnable code.
Build the orchestrator-workers agent pattern in Python with Claude: a planner that decides subtasks at runtime, parallel workers, and a synthesizer. Runnable code, July 2026.
A code-first routing tutorial: build a runnable LLM router in Python with the Claude API, add a deterministic fallback, and route by difficulty to cut cost.
Make a LangGraph agent durable with a checkpointer: attach SqliteSaver in one line, crash the graph on purpose, and resume from disk without re-running the Claude call.
A code-first prompt chaining tutorial: build a gated Python pipeline with the Claude API, add pure-Python and LLM gates, and fail fast before wasting tokens.
The exact claude_desktop_config.json setup to connect a local MCP server to Claude Desktop, verify it with the MCP Inspector, and fix a server that will not show up.
A minimal, framework-free LLM-as-a-judge harness in Python on Claude: rubric design, pointwise and pairwise scoring, and fixes for position, verbosity, and self-preference bias.
You cannot unit-test an agent like a pure function. Build a two-layer pytest harness: deterministic tool-call assertions plus an LLM-as-judge grader, a frozen eval dataset, and a CI gate. Runnable Python, no eval framework required.
