Going back to 2024, calling a system ‘agentic’ was a way to suggest it was sophisticated without committing to anything verifiable. The word covered everything from a chatbot with a function-call schema to a fully autonomous research system. It was useful exactly because it was vague.
That has slowly changed. The teams actually shipping these systems have converged on a tighter definition — one that is more useful in a procurement conversation and harder to fake in a demo. Here is the working version.
A working definition
A system is agentic to the degree that it (a) decides what to do next based on the state of its environment, (b) takes consequential actions in that environment, and (c) does both repeatedly without a human in the loop on every step.
All three properties matter. A system that deliberates beautifully but cannot act is a chatbot. A system that acts without deliberation is a script. A system that requires human approval on every step is, in the part that matters, a UI.
The interesting question is not whether a system is agentic. It is whether the agent is making decisions you would have made yourself, or decisions you wish you had thought to make.
Three failure modes worth memorizing
Most agent systems we see fail in one of three ways. None of them are exotic. All of them are predictable, and most are visible in a one-week pilot if you know to look for them.
- Tool soup. The agent has access to too many tools, picks the wrong one, and recovers slowly or not at all. Solvable with a tighter tool registry and better-designed tool descriptions, but the team usually wants to add tools, not remove them.
- Context rot. The agent’s context window fills with low-value history and the model starts making worse decisions late in a session than early. Solvable with explicit context-management strategies, but most starter frameworks ship with the wrong defaults.
- Confident wrongness in the long tail. The agent handles 90% of cases well, fails on the remaining 10%, and signals nothing different on the way down. The cost of the bad 10% often dwarfs the benefit of the good 90%.
A short evaluation checklist
Before you sign a contract for an agentic system — internal or vendor — ask the team to walk you through these. If they cannot answer cleanly, that is your answer.
- What does the agent do when its primary tool fails?
- How is its context managed across long sessions?
- What does the eval suite look like, and on what data was it built?
- Where is the human-in-the-loop boundary, and why there?
- How will you know when the model behind it changes underneath you?
None of this is novel. The teams getting agents into production have been asking these questions for over a year. The teams still struggling are usually the ones hoping a better model will let them skip the work that comes after the demo. It will not.