AI Guides › Skills & Agents

Giving An Agent A Budget, Not Just A Task

By Nigel Guy · 2 min read

A task tells an agent what to accomplish. It says nothing about how much that accomplishment is allowed to cost you — in time, in spend, in how long it runs before someone checks on it. Most agent setups skip this entirely, on the reasonable-sounding assumption that a well-scoped task will naturally stay within reasonable bounds. It often doesn't, particularly when the agent hits an unexpected obstacle and its response to "this isn't working" is simply to keep trying.

The mechanism

The Budget Card, alongside whatever task instructions the agent already has:

1. Set a time ceiling — a maximum number of steps, or a wall-clock limit, whichever the platform lets you measure. Even a rough number beats none; you can tighten it once you've seen a normal run's actual shape.
2. Set a cost ceiling — API spend, credits, whatever unit the underlying tools bill in. This matters most for anything that calls paid services in a loop, where a stuck agent doesn't just waste time, it spends money doing it.
3. Make hitting either ceiling a hard stop, not a soft warning. An agent that gets a warning and proceeds anyway hasn't actually been given a budget — it's been given a suggestion. The point of the ceiling is that crossing it pauses the agent and asks, rather than continuing on its own judgement.
4. Treat repeatedly hitting the ceiling as a design problem, not a reason to raise the number. If an agent keeps needing more time or spend than you budgeted, the honest fix is usually to understand why the task is taking that long, not to quietly widen the ceiling until it stops complaining.
5. Revisit the numbers as the task's real shape stabilises. Early ceilings are guesses; once you've watched several normal runs, set numbers based on what normal actually looks like, with headroom, rather than the original guess.

What to skip

Skip an open-ended budget on the assumption that "it's probably fine" — that assumption is exactly what a budget exists to stop you from relying on. And skip raising the ceiling as your first response the moment it's hit; find out why it was hit before deciding whether the number was wrong or the agent's behaviour was.

Guardrails

• A budget limits waste; it doesn't guarantee correctness. An agent can stay comfortably inside its budget and still produce a wrong result — the budget and the review habit are separate safeguards, not substitutes for each other.
• Repeated ceiling-hits are worth logging somewhere you'll actually see the pattern, not just handled one at a time as they come up — the pattern is often more informative than any single instance.
• Anything touching real spend deserves a cost ceiling checked against your actual tolerance for surprise charges, not a round number picked because it seemed generous enough at setup.

All 751 AI guides · JulieMango plans from £17/mo