Your agent doesn't know what it costs. Who's watching?

2026-10-08 — a coding agent that loops on a failing test doesn't feel the bill. It retries, spawns a helper, and retries again. Each step is locally reasonable. The total isn't. The money lives somewhere else: on your provider's billing plane, while the agent's process only sees requests going out.

Why spend limits are hard for agents

• The cost isn't in the process. Local tools can't see dollars at all, and the honest ones say so rather than inventing a number
• Provider caps are coarse. Account-level budgets and threshold alerts catch the month, not the runaway hour; by the time they fire, the work is done and billed
• A gateway sees the request, not the agent. It can meter traffic, but it can't tell which of the six agents on your machine started the loop, or which child process it spawned

What you can measure honestly today

• Count what the agent does, in units it cannot fake: runtime, billable requests, CPU time, memory held, LLM calls. Set limits in those units, and record every crossing
• A number without a source is “unknown”, not zero — an honest ledger beats a reassuring one
• That is what 20131's cost guard does today: local, open source, observation only. It records; it does not stop anything yet

A cap that survives its own outage

Every control depends on something that can fail. If your spend limit has to ask a remote endpoint and that call times out, gets rate-limited, or returns garbage, there are two lazy answers: allow everything, or block everything. Both are wrong. The contract we design against: on external-intelligence failure the runtime falls back to deterministic local policy, graded by the action's reversibility and blast radius — never binary. A budget cap that disappears with its own network calls is not a cap; it is a suggestion with latency.

The part that does not exist yet

Observation tells you what happened. A team running several agents wants a limit that holds the next action until a human says yes, and one view across everyone's machines. We are deciding whether to build that as a Team plan. It does not exist today, there is no price, and nothing on this blog gets ahead of the code.

If you run coding agents and a runaway invoice is the incident you would rather prevent than reconstruct, join the <a href="/en/early-access/">Early Access waitlist</a> or write to <a href="mailto:[email protected]">[email protected]</a> with the limit shape you would actually use. Every email is read by a human. If nobody asks for this, it stays unbuilt — that is the honest experiment behind this page.