What is an AI agent?

In practice

A useful agent is narrow. It has a small, explicit set of tools, a clear definition of done, a limit on how many steps it may take, and a log of everything it did. Agents that are given broad goals and open-ended tool access look impressive in a demo and behave unpredictably in production, which is why serious deployments keep the decision space small and the audit trail complete.

Where it fits in a build

Agents earn their place where the path genuinely varies case by case — triaging an incoming enquiry, chasing missing information across several systems, preparing a document for review. They are a poor fit for high-volume, fixed processes, where a deterministic workflow is cheaper, faster and easier to debug.

Common mistakes

• Giving an agent a tool that can send, pay, delete or publish without a human approving the action.
• Measuring the agent on whether it finished rather than on whether it was right, which hides failures until a customer finds them.
• Using an agent for something a simple rule would do, then paying model costs and latency for every single case.

Related terms

• Agentic AI
• Human-in-the-loop
• Workflow automation

Working on something that involves this?

ASTACKRA designs and builds AI systems, automation and custom software for businesses that need technology shaped around their own workflow. If this term turned up in a proposal and you want a straight answer about whether it applies to your situation, ask us — no obligation, and we will tell you if the answer is no.