Building Effective AI Agents: Three Ideas That Do Most of the Work
Anthropic's guidance on effective agents in three ideas: don't build agents for everything, keep them simple, and think like your agent. What each one means in practice.
Guillaume Rufenacht
AI Product Manager · Lisbon
The fastest way to waste money on AI is to build an agent for a job that didn’t need one. Anthropic’s own guidance on building effective agents comes down to three unglamorous ideas: don’t build agents for everything, keep them as simple as possible, and think like your agent. After shipping plenty of agents myself, I’d argue these three are most of the game.
Here’s what each one means in practice, and how I apply it to the production pipelines behind Geonimo.
Key takeaways
- Don't build agents for everything. Use a workflow when you can map the decision tree; reserve agents for ambiguous, high-value tasks.
- An agent is just a model using tools in a loop, defined by its environment, tools, and system prompt. Keep those three simple before optimizing.
- Think like your agent: everything it knows is the ~10-20k tokens in its context window. Most 'dumb' mistakes are missing context.
- Agents add agency, and with it cost, latency, and the consequences of errors. Earn that trade-off deliberately.
Don’t build agents for everything
An agent decides its own path; a workflow follows one you defined. Agents are how you scale complex, valuable, ambiguous tasks, not a drop-in upgrade for every feature. Before reaching for one, run the checklist:
Complexity
Value
Critical capabilities
Cost of error
Coding is the classic great fit: ambiguous (design doc to PR), genuinely valuable, and easily verifiable through tests and CI. That last property, verifiability, is doing a lot of work, and it’s exactly why I lean on the same discipline in my outbound system: a step that checks the work before it goes out.
Keep it as simple as possible
Strip away the mystique and an agent is a model using tools in a loop. Three components define it: the environment it operates in, the tools it can call, and the system prompt that sets its goals and constraints. Wildly different agents share almost the same backbone; the only real design decisions are which tools to offer and what to put in the prompt. Iterate on those three first, because complexity up front kills iteration speed. Optimizations, caching trajectories, parallelizing tool calls, come after the behavior is right. It’s the same lesson I cover from the other direction in tool, skill, or subagent?.
Think like your agent
When an agent does something baffling, the instinct is to call the model dumb. Usually the real problem is that we designed the agent from our perspective, not its own. Everything the model knows at each step is the roughly 10-20k tokens in its context window. Put yourself in that window: is the information actually sufficient and coherent to make the next decision? Do the exercise of running a task with only what the agent can see, it’s uncomfortable and clarifying.
The shortcut is that these systems speak our language, so you can ask Claude to understand Claude: paste your system prompt and ask if anything is ambiguous, or feed it a trajectory and ask why it made a given decision and what would have helped. It doesn’t replace your own understanding, but it closes the gap fast.
The takeaway
Reliable agents in production is the work I do. See how I think about context engineering, what I’ve built, or get in touch.
Frequently asked questions
When should I build an agent vs a workflow?
Use a workflow when you can map the decision tree explicitly; use an agent for genuinely ambiguous, high-value tasks where exploration justifies the token cost. Don't build agents for everything.
What is an agent, simply?
A model using tools in a loop, defined by three things: the environment it operates in, the tools it can call, and the system prompt setting its goals and constraints. Keep those simple before optimizing.
Why do my agents make confusing mistakes?
Usually because their context is insufficient. Put yourself in the agent's context window, everything it knows is the ~10-20k tokens it can see, and check whether that's actually enough to make the decision.
Work with me
Want a system like this built for your pipeline?
I help teams take AI from a clever prototype to dependable production, outbound engines, lead intelligence, and the LLM pipelines underneath. See what I have shipped or get in touch.