"Agent" has become one of the most overused words of the moment, often used to describe any automation that involves an AI model. It's worth being precise: an agent is a system that plans a sequence of actions, calls external tools, evaluates the results, and decides the next steps — not a single prompt with a single answer.
Where open models work well today
- Tool calling against well-defined, documented APIs, with a limited number of possible actions.
- Research and synthesis agents over company documents (agentic RAG), where the task is to retrieve, compare, and summarize existing information.
- Automation of workflows with known, predictable steps — classification, routing, guided form-filling — with clear guardrails on what the system can and cannot do.
- Structured document processing (data extraction, validation, pre-filling) with a final human check.
Where caution is still warranted
- Autonomous planning over long horizons, with many sequential steps and little supervision: the margin for error grows with every step.
- Ambiguous instructions that require common sense or tacit knowledge of the company context that isn't written down anywhere.
- High-impact, hard-to-reverse actions (payments, external communications, changes to production systems) without a human confirmation step.
The principle we apply with clients
We start from a narrow, well-defined scope, with a person in the loop for decisions that matter, and expand the agent's autonomy only after gathering evidence that the system behaves reliably on that specific task. It's a less spectacular approach than a demo that does everything on its own, but it's the one that holds up when the agent has to work every day, not just in a presentation.