How We Build Effective Agents

Insights from Barry Zhang (Anthropic)

A battle-tested playbook for shipping reliable AI systems.

How We Got Here

Phase 1: Simple Features — Summarization, classification, and basic extraction. (Now baseline table stakes).
Phase 2: Workflows — Orchestrating multiple model calls in predefined, hardcoded control flows. Trades latency/cost for reliability.
Phase 3: Agents — Systems that can evaluate environment feedback, dynamically decide their own paths, and operate independently.

1. Don't Build an Agent. Build a Workflow.

If you can map out the decision tree, a workflow wins every time.
Workflows are cheaper, faster, and more reliable.
Agents are reserved for ambiguous, high-value tasks only.

Use a Workflow if:

You need complete node-by-node control, predictability, and have a strict token budget (e.g., under 10 cents/run).

Use an Agent if:

The trajectory cannot be hardcoded, outcomes are highly variable, and the output value easily justifies high token consumption.

2. Keep it Embarrassingly Simple.

An effective agent consists of exactly 3 fundamental things:
Environment + Tools + System Prompt
That's it. Caching, parallelization, orchestration, and multi-agent systems are entirely downstream.
Any unnecessary complexity introduced up front completely kills your iteration speed.

3. Think Like Your Agent, Not Like Yourself.

Your agent only sees its immediate 10k–20k token prompt, not your massive human context window.
It actually reads the rushed tool descriptions you wrote at 11:00 PM. Be specific.
Operate like the agent: realize that executing actions on a UI blindly without instant feedback is like coding completely in the dark.
The Audit Loop: Paste your final system prompt right back into Claude. Ask it what parts are ambiguous, and fix exactly those.

The Core Realization

                "Most builders don't actually have an agent problem. 

                They have a workflow problem dressed up as an agent."

Autonomy isn't the final goal. Reliability is.

The Next Engineering Frontiers

Budget-Aware Design

Enforcing strict, native execution caps on financial costs, time execution, and total tokens used before an agent spins out of control.

Self-Evolving Tooling

Leveraging a meta-tool layer where agents evaluate and rewording their own tool descriptions for better internal ergonomics.

Asynchronous Multi-Agent Systems

Transitioning from rigid, synchronous User-Assistant turns to lightweight sub-agents communicating across parallel boundaries to save core context space.

Save This Before Your Next Build

1. Don't build an agent. Build a workflow. (Map it if you can).

2. Keep it embarrassingly simple. (Environment + Tools + Prompt).

3. Think like your agent. (Debug its actual token window perspective).

                The teams shipping working AI in 2026 won't be the ones with the cleverest architecture... 

                They'll be the ones who picked the boring one and shipped twice as fast.