Anthropic (Engineering blog) · Guide
Building Effective AI Agents
The engineering baseline everyone borrows: start simple, prefer workflows, and invest in tools more than prompts.
- Published
- Reading time
- About 15 min
- Authors
- Erik Schluntz, Barry Zhang
The live page has been edited since 2024; quote the edition you mean.
What is it about?
Drawing on work with dozens of teams, Anthropic's engineers found that the most successful agent implementations used simple, composable patterns rather than complex frameworks. They distinguish workflows, where LLMs and tools follow predefined code paths, from agents, where the model directs its own process and tool use.
The guide lays out a ladder: a single optimized LLM call with retrieval, then workflow patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer), and only then autonomous agents for open-ended problems. It warns that the right answer “might mean not building agentic systems at all”.
Key numbers
- While building its SWE-bench coding agent, Anthropic spent more time optimizing tools than the overall prompt.
- Switching from relative to absolute file paths made the model use its tool “flawlessly”.
- Two domains where agents proved valuable: customer support and coding, both with clear success criteria.
Why read it
- It gives a CIO shared vocabulary with engineers (workflow versus agent) and a defensible default: add complexity only when it measurably improves outcomes.
- The section on the agent-computer interface, which says to document tools as you would for a junior developer, is where many teams under-invest.
- It is short enough to read in one sitting and is still the reference most agent guides cite.
Who should read it
- CIOs and engineering leaders setting agent architecture standards
- Product managers scoping “agent” features
- Non-technical executives who want to know what an agent actually is
About the author: Erik Schluntz
Member of technical staff, Applied AI, Anthropic
Before Anthropic, Schluntz co-founded Cobalt Robotics, which built security robots for offices, and served as its CTO. When integrating with each elevator company's API took around six months, he added a simple arm that pressed elevator buttons using computer vision: deployable in a day, if slower and less reliable.
Earlier he took time off from Harvard to go through Y Combinator, worked on Google[x]'s smart contact lens, and led testing for a Falcon 9 fuel-system software overhaul on SpaceX's flight software test team. At Anthropic he built the SWE-bench agent behind this guide's best anecdote.
Read it carefully: common misquotes
- It does not say “don't build agents”. It says you may not need them.
- Benchmark numbers often attached to it come from later Anthropic posts.
How it shapes the CIO's AI Playbook
- Step 3.3: Use the simplest architecture that works · Redesign the work, not the task
- Step 3.4: Invest in tools more than prompts · Redesign the work, not the task
Notes by Pavel Veselov, last reviewed . Summaries are my own; quotes and figures belong to the original authors.
All 11 sources in the CIO's AI Library