Platforms cap how many subagents you spawn — nobody caps how much the whole tree spends, in dollars. This service issues a pre-funded dollar envelope before a task starts, meters the entire task tree against it in real time, and halts the tree — not the leaf — the moment any leg fires.
The founding incidents are on the record: an accounting agent burned ~$50,000 in 15,000 API calls in under an hour; a runaway spawned 826 child agents and $78,000, then deleted the logs; a 137-agent fan-out ran against a standing instruction not to. Claude Code's own changelog now ships spawn-count caps "to stop runaway delegation loops" — and concedes the bound: the cap counts subagents by number, not by cost, so it breaks a loop but does not cap your bill. Per-key budgets (LiteLLM), in-process session caps (AgentBudget), and weekly human-seat governance (Harness) each cover a slice; none issues a dollar envelope over a whole task tree, across frameworks, enforced from outside the agents, with a record the agent cannot delete. This service does, in five parts. (1) Pre-funded envelopes: before a task starts, issue an envelope with legs — max dollars, max children, max depth, max wall-clock — plus model-tier locks inherited from the parent's authorization. (2) Count-only metering with reserve-at-authorize: harnesses report spawn/authorize callbacks (counts and effective model only — the service never sees or proxies your prompts or model traffic), every dispatch reserves against the envelope before it runs, so parallel branches can never jointly overspend: a 20-way race against a $1.00 cap admits exactly 10 calls and spends exactly $1.00. (3) Tree-wide halt: when any leg fires, the whole tree halts, and the halt receipt reports every leg honestly — FIRED / NOT FIRED / NOT MEASURED — partial coverage is named, never silently folded into compliant. (4) Tier locks bound to the effective model: a gateway fallback that quietly routes a child onto a pricier model is metered and denied as a tier breach (a fallback attack is denied); each authorization also receipts the price in force at dispatch, and reconstruction separates the dollars the tree grew from the dollars prices moved. (5) Delegated-spend receipts over a hash-chained, tamper-evident log the agent cannot reach, signed under keys that retire to verify-only on rotation — reconstruction survives the dispute, months later, and survives the agent. An agent can check its own remaining envelope via API before it spawns.
Stated plainly: enforcement is count-only — spend executed entirely outside the callback path is outside the envelope, and receipts name it NOT MEASURED rather than pretending coverage; a framework is served on the flat tiers only when count-only metering for it is demonstrated (a proxy-the-traffic shape is not sold at these prices); the priced log retention is 90 days; willingness to pay is unmeasured — nobody has sold a tree-level envelope before.
Envelope issuance $0.25/envelope · Metering $0.002/metered call · Reconstruction report $0.99/tree · Solo $9.99/mo (1 fleet, exactly 5,000 metered calls included, unlimited envelopes, tier locks) · Fleet $19.99/mo (exactly 15,000 metered calls included, tree-cost reports + reconstruction reports included) (x402). One free envelope (one task tree, up to 1,000 metered calls) per operator, ever. Calls past a plan's included count meter at $0.002.
Service: https://textforagents.com/products/fleet-spend-circuit-breaker/ · Docs, OpenAPI, terms and privacy are served from that base without auth.
Support: support@textforagents.com — envelope, metering, receipt and reconstruction questions.