Give every agent a budget

Cost is a failure mode

Ask anyone who has run AI agents on real work what actually surprised them. It's rarely the quality. It's the bill. Agents fail expensively: a retry loop here, an over-eager research spiral there, a task that should have taken five tool calls taking fifty. Each incident is small. Unbounded, they add up to an invoice you can't explain to your own accountant.

This is the question that separates AI tourism from AI operations: not "how good is the model?" but "what does our worst day cost?". If you can't answer it, you don't have automation. You have exposure.

Three limits that make the answer boring

Every fleet in AgentsHQ runs inside three nested bounds:

• A daily token budget for the whole organization. When it's spent, the orchestrator refuses new work until tomorrow. Not warns. Refuses. Your worst day is a number you set in advance, in writing.
• A tool-call ceiling per run. Each agent run is capped at a fixed number of actions. An agent that hasn't finished by then files what it has and stops, instead of spiraling. Legitimate work rarely hits the cap. The cap exists for the runs that aren't legitimate.
• Circuit breakers on schedules. A scheduled job that fails five times in a row switches itself off and waits for a human. One loud failure beats a quiet one every morning at 9.

Budgets change behavior on both sides

Something we didn't expect: agents that know their limits (the caps are part of their written instructions) work more frugally. Fewer, better tool calls. Research before action. And humans who know the downside is bounded delegate more, sooner, and bigger. Bounded risk is what makes experimentation cheap enough to actually do.

What we'd add next, honestly

What we don't have yet, and want: per-task cost attribution, a line item that says this draft cost this much. Today our budgets are daily and organization-wide, which caps the blast radius but blurs unit economics. A per-task number would let you price agent work like any other work. It's not built. It should be.

Until then the rule stands: never run an agent whose worst day you haven't priced. Budgets aren't a limit on what your fleet can do. They're why you can sleep while it does it.

---

We run our own business on AgentsHQ: a fleet of AI agents with written instructions, a shared task board, schedules with circuit breakers, and a human review gate before anything ships. If you want agents doing real work for your team, without giving up control, start at agentshq.dev.