AI agent
An AI agent is a system that, given a goal, plans a sequence of steps and uses tools — searching, calling APIs, writing files — to accomplish it with little or no human intervention. Unlike a single prompt-and-answer, an agent loops: it acts, observes the result, and decides the next step.
The core of an agent is a loop around an LLM: the model is given a goal and a set of tools, it chooses an action, the system runs it, and the result is fed back so the model can choose again until the goal is met. This is what lets software book the cheaper route, dedupe a feed, or triage a queue without a person directing each move. Reliability comes from the scaffolding — clear tool definitions, retries, state, and stop conditions — far more than from the model alone.
When building a product, the honest question is whether the task genuinely needs an agent or whether a fixed workflow with one or two model calls would be more reliable and cheaper. Agents shine when the path can't be predicted in advance; they are risky when each step can fail silently and compound. We build them with explicit guardrails, evals on the whole trajectory (not just the final answer), and a human checkpoint wherever a wrong action would be expensive — our worldgovdata ingestion pipeline, for example, runs unattended across 164 sources but gates every publish behind contract assertions — when a check fails the last good data stays live, rather than the loop being trusted blindly.
// faq
Frequently asked questions
- When should I use an AI agent instead of a simple workflow?
- Use an agent when the steps can't be known ahead of time and depend on what earlier steps return — research, triage, multi-tool problem solving. If the steps are fixed and predictable, a plain workflow with one or two model calls is cheaper, faster, and far easier to test. Don't reach for an agent because it sounds advanced; reach for it when the task is genuinely open-ended.
- Why do AI agents fail in production?
- Usually compounding errors and weak observability. Each step has a small failure chance, and over a long loop those multiply, so one bad tool call derails everything downstream. The fixes are engineering, not a bigger model: tight tool definitions, retries and timeouts, state you can inspect, trajectory-level evals, and human checkpoints on costly actions.