AI agent

An AI agent is an LLM-backed system that determines, at runtime, what to do next - which tool to call, what information to gather, whether to finish or iterate. The contrast is a workflow automation, whose path is knowable in advance.

The decision rule

• If you can write the workflow as a flowchart in ten minutes, it's probably an automation with a model call at a decision point.
• If you can't, it's probably an agent.

Getting this wrong is the single most common architectural mistake in production AI systems - we walk through the full framework in AI agents vs automation.

What agents are good at

• Multi-step research and synthesis tasks.
• Triage of novel cases where the right next step depends on what's discovered.
• Drafting tasks that require gathering context from several tools.

What agents are expensive at

• High-volume tasks where an automation with one model call would suffice. Agents typically make 5–20 model calls per invocation.
• Tasks with strict deterministic requirements.
• Systems without a strong evaluation harness
  • agents fail in interesting ways that casual monitoring misses.

Governance implications

Agents are harder to govern than automations. The logic is partly in the prompt, partly in the tool definitions, and partly in the runtime decisions of the model. An audit trail for an agent has to record every tool call, not just the end-to-end input and output.