Anthropic Shares Three Design Patterns for Building Better AI Agents with Claude
Key Takeaways
- ▸Agent harnesses should encode fewer assumptions and strip down as models improve; design patterns should be revisited as Claude's capabilities evolve
- ▸Providing Claude with general tools (bash, text editors, code execution) and letting it orchestrate results reduces latency and cost compared to processing every tool output through the context window
- ▸Strong coding models are strong agents; effective agent design centers on what the model can already do rather than compensating for perceived limitations
Summary
Anthropic published guidance on agent harness design—the software scaffolding around AI models—outlining three patterns for building applications that balance intelligence, latency, and cost while keeping pace with Claude's evolving capabilities. The core insight is that agent harnesses encode assumptions about what Claude can't do on its own, and those assumptions become stale as the model improves. Anthropic's three recommended patterns are: leveraging what Claude already knows well (such as bash and text editors), progressively stripping down the harness as Claude becomes more capable, and carefully setting boundaries for the agent.
The article draws on concrete examples from Claude's development, including Claude 3.5 Sonnet's 49% performance on SWE-bench Verified using only bash and text editor tools. Anthropic emphasizes that giving Claude code execution capabilities allows it to orchestrate its own actions and filter tool results without unnecessary token overhead, moving orchestration decisions from the harness to the model itself. This approach has proven effective—equipping Opus 4.6 with output filtering improved web browsing accuracy from 45.3% to 61.6%.
Editorial Opinion
This framework reflects pragmatic maturity in AI application design: rather than over-engineering scaffolding around models, teams should design for transparency and let improving models handle more decisions themselves. The emphasis on stripping away assumptions rather than adding guardrails runs counter to common instinct but aligns with how capability actually advances—by expanding what the model can do reliably on its own. As Claude and other frontier models continue improving, this principle will likely become table stakes for competitive agent applications.
