Building Effective AI Agents

Source: Anthropic Engineering Date: 2024-12-19 URL: https://www.anthropic.com/engineering/building-effective-agents

Summary

Anthropic’s canonical agent architecture reference, describing six patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer, autonomous agents) with a strong bias toward simplicity over complexity. A concrete example from their SWE-bench work: more time was spent optimizing tools than the overall prompt, and requiring absolute file paths in tool interfaces eliminated model mistakes with relative paths.

Implications

The foundational harness design reference. This is the post all subsequent Anthropic harness content builds on — the evaluator-optimizer pattern described here becomes the three-agent harness in the 2026 long-running apps post. Reading it alongside those later posts shows how the patterns have evolved from taxonomy to concrete production implementations.

ACI over prompting. The emphasis on Agent-Computer Interface design — tool documentation, interface constraints like absolute paths — over prompt tuning is the specific insight that separates Anthropic’s published guidance from generic LLM best-practices content. It directly informs tool design for Claude Code integrations.

Frameworks vs. bare components. The explicit recommendation against framework lock-in for production systems is a positioning move that favors Anthropic’s own primitives (Claude API + MCP) over LangChain, CrewAI, et al. This is the document that validated the “build with basics” approach for shops running production agents.