How Anthropic Builds Effective Agents: Four Questions to Answer First

Barry Zhang from Anthropic took the stage at the AI Engineer Summit to share how the team behind Claude thinks about building effective agents. Drawing from Anthropic's widely-read blog post "Building Effective Agents," he presented three core ideas and a practical checklist for deciding when agents are the right approach.

The Evolution of Agentic Systems

The journey to agents has been incremental. It started with single LLM features -- summarization, classification, extraction -- tasks that felt like magic two to three years ago and are now table stakes. As products matured, developers started orchestrating multiple model calls in predefined control flows, trading cost and latency for better performance. Anthropic calls these "workflows."

Agents represent the next step. Unlike workflows, agents can decide their own trajectory and operate almost independently based on environment feedback. But with more agency comes more cost, more latency, and higher consequences when things go wrong.

Idea 1: Don't Build Agents for Everything

Agents are a way to scale complex and valuable tasks. They should not be a drop-in upgrade for every use case. Here is the checklist Anthropic uses to evaluate whether an agent is the right approach:

Is the task complex enough? Agents thrive in ambiguous problem spaces. If you can map out the entire decision tree easily, just build a workflow and optimize every node. It is more cost-effective and gives you more control.

Is the task valuable enough? Agent exploration costs tokens. If your budget per task is around 10 cents -- say, a high-volume customer support system -- that only affords you 30,000 to 50,000 tokens. Use a workflow instead.

Are all parts of the task doable? De-risk the critical capabilities. If you are building a coding agent, make sure it can write good code, debug, and recover from errors. Bottlenecks will multiply your cost and latency.

What is the cost of error and error discovery? If errors are high-stakes and hard to discover, it becomes very difficult to trust the agent to act autonomously. You can mitigate with read-only access and human-in-the-loop, but that limits scalability.

Idea 2: Keep It Simple

Agents are models using tools in a loop. Three components define what an agent looks like: the environment (the system the agent operates in), the tools (the interface for taking action and getting feedback), and the system prompt (goals, constraints, and ideal behavior).

The Anthropic team has learned the hard way to keep this simple. Any complexity introduced up front kills iteration speed. Iterating on just these three basic components gives by far the highest ROI. Optimizations -- caching trajectories, parallelizing tool calls, presenting progress to build user trust -- can come later.

Idea 3: Think Like Your Agent

Developers often build agents from their own perspective and get confused when agents make mistakes. The recommendation is to put yourself in the agent's context window. At each step, the model is running inference on a limited set of context -- typically 10,000 to 20,000 tokens. If that context is not sufficient and coherent, the agent will struggle.

For computer use agents, this means understanding that the agent only sees a static screenshot, executes a click without visual feedback, waits three to five seconds in the dark, and then sees another screenshot. Trying to complete a task from this perspective makes it immediately obvious what context the agent actually needs.

Looking Ahead

Three areas that occupy Anthropic's thinking: making agents budget-aware (defining and enforcing limits on time, money, and tokens), enabling self-evolving tools (agents that can design and improve their own tool interfaces), and multi-agent collaboration in production (with the open question of how agents communicate asynchronously).

Key Takeaways

• Don't build agents for everything -- use the four-question checklist: complexity, value, capability bottlenecks, and error cost
• Keep agents simple: environment + tools + system prompt, then iterate
• Think like your agent -- put yourself in the context window and ask if the information is sufficient
• Coding is an ideal agent use case because the output is verifiable through tests and CI

Resources

• Building Effective Agents (Anthropic Blog) -- The original blog post by Barry Zhang and Eric
• AI Engineer Summit -- Conference where this talk was delivered