← back
How We Build Effective Agents: Barry Zhang, Anthropic
Takeaway
Use agents only when tasks are ambiguous, valuable, and reversible, keep the architecture minimal, and debug by stepping inside the model's context.
Summary
- Anthropic's Barry Zhang revisits the 'Building Effective Agents' blog post and argues three principles: don't build agents for everything, keep them simple, and think like your agents.
- Agents should be reserved for complex, ambiguous, high-value tasks where errors are tolerable; workflows with explicit decision trees are cheaper and more controllable.
- Simplicity centers on three components: a model, tools/environment, and a feedback loop — adding control flow before it's needed hurts reliability.
- Engineers should debug by reading full agent trajectories with the agent's limited context window in mind, rather than only looking at end outputs.
agentsanthropicworkflows
Original description
Recorded live at the Agent Engineering Session Day from the AI Engineer Summit 2025 in New York. Learn more at https://ai.engineer and purchase tickets to our next event, the AI Engineer World's Fair, in SF June 3 - 5 here: https://ti.to/software-3/ai-engineer-worlds-fair-2025 About Barry: Barry is a member of technical staff on Anthropic's Applied AI team, focusing on developing agentic systems with enterprises and startups. Previously, he was a tech lead on the Monetization genAI team at Meta, where he claimed the inaugural 'AI Engineer' title. He holds degrees in Computer Science and Industrial Engineering from Northwestern