When people say “we deployed an agent,” (opens in a new tab) they usually mean “we connected a model to some tools.” That gap is where the horror stories live: the $47,000 runaway invoice (opens in a new tab) one developer wrote up, the half-billion-dollar month (opens in a new tab) one consultant described to Axios. An agent that fails expensively is almost always an agent that was missing one of its parts.
An agent is six things, not one:
The six parts
- Model. The reasoning engine. The part everyone talks about, and the least differentiated. Models are increasingly a commodity.
- Memory. What it knows and remembers across steps. Without it, every action starts from zero.
- Hands. The tools and actions it can take in the real world: send the email, move the money, update the record. Hands are where value and danger both live.
- Clock. Its sense of when: what runs on a schedule, how long a task may run, when to stop. Missing clocks are how you get the runaway bill (opens in a new tab).
- Gates. The checks: what requires approval (opens in a new tab), what it may never do, where a human must sign off. Gates are governance made concrete.
- Briefing. The instructions and context that tell it what good looks like (opens in a new tab). A vague briefing is a confident agent doing the wrong thing well.
Why this matters
Gartner has warned (opens in a new tab) that a large share of agentic AI projects, on the order of 40%, will be cancelled by 2027, and cost overruns are a leading reason. The pattern is consistent: teams invest in the model (part one) and under-invest in the clock and gates (parts four and five). The result is capability without control.
What to take into the room: before you approve an agent, ask for all six parts by name. If someone can describe the model and the hands but goes quiet on the clock and the gates, you are looking at a budget incident waiting to happen.