Long-running jobs
Remote and managed agents can keep working after the first prompt. The review checks where the system stops by policy, not by trust.
tate@programs
~/services/agent-spend-guard
caps / retries / bills / paid tools
agent spend guard / may 2026
Before agents run longer, set the hard stop.
Coding agents, managed agents, browser agents, and paid API calls can now run in loops across files, tools, payments, and background jobs. This review maps the spend blast radius and the controls that stop a small task from becoming a real bill.
- entry
- $99 quick map
- full review
- $199
- delivery
- 48h
why now
Agent infrastructure moved faster than most billing controls.
Tool fleets
One task may call browsers, file tools, paid APIs, search, model endpoints, webhooks, and queues. The map shows every spend path.
Retry storms
Failed jobs, partial payments, transient 500s, and invalid tool output can multiply cost unless retries have budgets and idempotency.
review checklist
What gets checked.
The review stays inside the authorized project, docs, logs, repo, or sandbox you provide. It does not place trades, move funds, create accounts, bypass limits, or trigger paid calls unless you provide a written test scope.
Models, paid APIs, x402 routes, browser actions, job queues, scheduled runs, hosted tools, storage, and webhook fanout.
Per-call, per-task, per-user, per-session, per-day, and per-provider limits enforced outside prompt text.
Duplicate request ids, exponential backoff, max retries, paid-but-denied recovery, queue poison messages, and cron re-entry.
Thresholds where the agent must stop for human approval before spend, account changes, deployments, or external publication.
Usage logs, denial logs, payment receipts, provider bills, trace ids, and reconciliation fields that prove the cap actually worked.
Immediate stop path for one user, one project, one key, one provider, or the whole agent fleet without waiting on code deploys.
deliverable
A short private patch order, not a giant audit.
map
Cost surface diagram.
Every route from user action to model/tool/API spend.
caps
Hard-stop table.
Where caps live, what they block, and what evidence proves it.
patch
Smallest next fixes.
Prioritized by real billing blast radius and implementation effort.
source trail
Why this service exists now.
Managed agents and Antigravity
Google's 2026 developer announcements moved agent infrastructure closer to managed, production-facing workflows.
developer keynote notesCodex-style work is operational
Agent work increasingly spans files, tools, review loops, and remote execution, which makes cost boundaries part of launch readiness.
Codex appSDK and MCP plumbing shifted
The Anthropic/Stainless acquisition made generated clients and tool interfaces a board-level infrastructure topic.
TechCrunch reportRunaway usage is visible
Large public agent-runtime bills made token and tool spend governance an urgent product concern, not a theoretical one.
usage storyStart here