Sources checked
What is an AI agent?
In Building effective agents, published 19 December 2024, Anthropic groups both patterns under “agentic systems” and then splits them. Workflows orchestrate models and tools through predefined code paths.
Agents let the model direct its own process and tool use.
The essay says the augmented building block is a model with retrieval, tools, and memory. On top of that, the authors describe prompt chaining, routing, parallel work, an orchestrator with workers, and an evaluator that critiques a draft.
Those are still workflows when the path is fixed. The agent pattern is a loop: look at the tool result, choose the next action, and stop when the task is done or a limit is hit.
How do you create an AI agent?
Anthropic’s own examples are a coding agent on SWE-bench tasks and a computer-use reference implementation. Both can be checked from the outside.
Their three principles are simple design, visible planning steps, and a carefully specified tool interface. Our guide to evaluating an AI agent is the scorecard for that check, and the Model Context Protocol is one way tools are connected.
The protocol does not decide whether the agent should be trusted.
- Name the finished result in a way you can observe.
- Give only the tools that job needs, with inputs the model can fill correctly.
- Show the plan and the tool results, not only the final message.
- Stop after a set number of steps, or when a person must approve an action.
Agent vs workflow: what is the difference?
The same broken page can be handled either way. The difference is who chooses the next step.
| Fixed workflow | Agent-style process |
|---|---|
| Always read a log, summarize it, and return a checklist. | Inspect the page, choose relevant files or checks, and adapt to the result. |
| The sequence is set in advance. | The latest tool result helps determine the next action. |
When should you not build an agent?
Anthropic says agentic systems trade latency and cost for flexibility, and that many applications only need one well-written model call with the right context.
Use an agent when you cannot predict the number of steps and you can tolerate the model choosing them. If the path is known, a workflow is the more predictable system.
Common questions
Is every chatbot an agent?
No. A conversation box does not show that the system chooses actions or uses tools.
What is a local AI agent?
It is an agent whose model and tools run on hardware you operate, rather than a vendor’s hosted loop. The same checks apply.
Where it runs is covered in the local AI agents guide.
Does more autonomy guarantee better results?
No. Anthropic warns that errors can compound across steps. Judge the completed work.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards





