Agent
An AI system that can autonomously plan, execute and iterate on a multi-step task using a language model plus tools. Different from a chatbot, agents act, they don't just respond.
An agent is an AI system that can autonomously work toward a goal by chaining together model decisions, tool calls and observations. The key difference from a chatbot is the loop: an agent decides what to do, does it, observes the result, decides what to do next, and continues until the task is complete.
The simplest agent is: a language model + the ability to call tools + a loop that runs until the model says “I’m done.” Anthropic’s reference example is Claude Code itself, it can read files, edit code, run shell commands, observe the results, and keep going.
What makes something an agent vs a chatbot
| Chatbot | Agent |
|---|---|
| One turn at a time | Many turns toward a goal |
| User chooses each step | System chooses each step |
| No tools (usually) | Has tools (web, file system, APIs) |
| Stateless or short context | Maintains state across many actions |
Common agent shapes
- Single-task agent. “Audit my Shopify inventory tonight and post a summary.” Runs to completion, exits.
- Conversational agent with tools. Claude.ai with web search + file uploads. Mostly chat-shaped, but can take agentic actions when needed.
- Multi-agent. A coordinator agent that dispatches work to specialist subagents (research, code, write). Each subagent runs its own loop and returns a result.
- Background agent. Runs on a schedule (cron, GitHub Actions) without a human in the loop. Common for monitoring, scraping, reporting.
Risks specific to agents
Because agents act autonomously, the failure mode is “wrong action, executed.” This is more expensive than a chatbot’s “wrong sentence, displayed.”
Mitigations:
- Scope tools narrowly. Only the tools needed for the specific task.
- Permissions. Read-only first; promote to write only after observing behaviour.
- Budget caps. Token caps, turn caps, time caps.
- Audit logs. Record every tool call.
- Human approval gates on consequential actions (sending emails, posting to social, writing to production systems).
Modern agent frameworks (Claude Code, OpenAI Agents SDK, Anthropic’s Claude Agent SDK) build these mitigations in. Use them.
Related terms
Get the next one in your inbox
Want this built for your business?
Book a free 30-minute AI audit. We'll map your business and show you exactly which systems we'd build first. No pitch deck, no scoping fee.
Book my free AI audit