AI agents and tool use

An AI agent is a system where a language model can plan, use tools, observe results and decide what to do next — in a loop. Unlike a simple chatbot that receives a message and returns a response, an agent can break a task into steps, execute actions (search the web, write code, query a database), observe the results, and decide whether to continue, retry or stop.

The key difference from a chatbot is autonomy: an agent takes multiple actions to accomplish a goal, rather than just answering a question.

Every agent follows the same core loop: think, act, observe, repeat.

Think: the model receives the current state and decides what to do next. Act: the model calls a tool (search, code execution, API call). Observe: the model reads the result of the tool call. Repeat: based on the observation, the model decides whether to call another tool, retry, or return a final answer.

This loop continues until the agent decides the task is complete, hits a maximum number of steps, or encounters an error it cannot recover from.

Agent demos look impressive. Agent products mostly struggle. The core problem is compounding errors: if each step has a 90% success rate and the task requires 10 steps, overall success is only 35%. Real-world agents hit ambiguous tool outputs, unexpected errors, and decision points where the model makes the wrong choice.

Other failure modes include: runaway loops (the agent keeps retrying without making progress), excessive tool calls (wasting time and money), and security issues (an agent with too much access can cause real damage).

The companies shipping successful agents constrain them heavily — limited tools, short loops, human approval for risky actions, and extensive eval suites.

Agents work best on well-defined, bounded tasks with clear success criteria. Code assistants (Claude Code, Cursor, GitHub Copilot) are agents that read code, make changes and run tests. Customer support agents handle tickets by searching knowledge bases and taking actions. Data analysis agents run SQL queries and create visualisations.

The pattern that works: narrow scope, good tools, short loops, human oversight. The pattern that fails: 'autonomous agent that can do anything'.

You can build agents with frameworks like LangChain, CrewAI, or Anthropic's agent SDK — or you can build them from scratch with a simple while loop and tool-calling API. Start simple. Build an agent that does one thing well (e.g., research a topic using web search) before trying to build a general-purpose assistant.

The most important skill is designing good tools: clear names, typed parameters, helpful error messages, and bounded scope. The model is only as good as the tools you give it.