Agents fail silently
When an API call fails, most agents retry. Without a circuit breaker, retries compound costs. Without alerts, nobody knows until the bill arrives.
Meta's rogue agent incident: unauthorized access, zero alerts
Born from a $10 incident where uncontrolled agents burned budget silently for 37 minutes. OpenClaw is an open-source agent harness with context compaction, spending guards, and circuit breakers.
It started with an overnight agent run. No spending limits. No circuit breakers. No alerts. The agent ran for 37 minutes, made 14 API calls, and burned $10 before anyone noticed.
An autonomous agent was left running with no cost ceiling. It entered a retry loop on a failing API call, each attempt consuming tokens at Opus pricing. No alert fired. No circuit broke. The agent kept going — silent, expensive, and utterly pointless. $10 gone in 37 minutes on zero useful output.
The money wasn't the problem. The silence was. An agent that can spend without limits, fail without tripping, and run without oversight is not autonomous — it's uncontrolled. OpenClaw was built so that never happens again.
As AI agents gain autonomy — running multi-step workflows, making API calls, managing files — the failure modes shift from "wrong answer" to "uncontrolled spending." The industry is building agent capability. Almost nobody is building agent governance.
When an API call fails, most agents retry. Without a circuit breaker, retries compound costs. Without alerts, nobody knows until the bill arrives.
A 200k-token Opus session costs $3-8. Most of that context is stale conversation history. Without compaction, you pay full price for information the model already summarized three turns ago.
Enterprise adoption of agentic AI requires predictable costs. You cannot deploy autonomous agents to production without spending limits, approval flows, and audit trails.
Today it's Claude. Tomorrow it might be Gemini or an open-source model. The harness layer must treat models as interchangeable compute — route to the cheapest model that can handle each task.
Every API call passes through three checkpoints. Each layer can stop the operation independently. Together they make runaway spending structurally impossible.
Estimates cost before execution. Checks against daily, weekly, and monthly limits. Operations above $2 require explicit approval. Budget depletion is blocked outright.
Wraps every API call. After 2 consecutive failures, the breaker trips — rejecting all calls immediately. After 60s cooldown, allows one test call. Recovers on success, re-trips on failure.
Logs actual usage after every call. Evaluates alert thresholds. Produces daily cost reports with model-level breakdown, forecasting, and trend analysis.
Monitors token count per session. At 170k tokens, automatically summarizes old turns via Haiku, preserving the 20 most recent. Extracts learnings to MEMORY.md for cross-session persistence.
Every layer is opt-out, not opt-in. OpenClaw ships with safe defaults — $2/day limit, 2-failure circuit breaker, 170k auto-compact. You relax constraints deliberately, not accidentally. The $10 incident happened because there were no defaults at all.
Every operation gets one of three verdicts: ALLOW (proceed silently), APPROVE (ask the human), or BLOCK (refuse outright). No fourth option.
OpenAI, Anthropic, Google, and Meta are building walled gardens. The cost control and governance layer cannot be owned by the same companies selling the tokens. It must be open infrastructure.
If your spending controls live inside the provider's SDK, switching models means rebuilding your governance layer from scratch. OpenClaw sits outside any provider.
The $10 incident happened because there was no library for this. Now there is. The next person who deploys an autonomous agent doesn't need to learn the same lesson.
Without spending guards, only organizations with large budgets can safely experiment with agentic AI. With a $2/day default, anyone can run autonomous agents without fear.
# Install npm install openclaw # Start monitoring immediately openclaw cost monitor # Check your guard status openclaw guard status # View spending dashboard openclaw cost report
"Go back and get it." The Adinkra principle that drives OpenClaw. Every mistake becomes infrastructure. The $10 incident became a library. The library became a thesis. The next agent you deploy will never burn budget silently.