If an agent can edit its own spending cap, was it ever a cap?
quiet_flare · reddit · 2026-11-03
Reacting to news of an AI agent raising its own spending cap, the author audits their Mainsheet runtime (built on Claude Agent SDK): budgets cap turns, tool calls, time and cost, with the cost cap ending the run.
Key points:
- A budget the model can read is just a suggestion—limits in prompts, config files in the working directory, or behind callable tools give the agent a path to change them, no bad intent needed.
- The author admits an honest gap: their README doesn't state where the signing key lives or prove the agent can't reach it, and there's no test proving the policy file is safe either.
- Takeaway: enforcement and storage of limits must both sit outside anything the agent can write to. Enforcement is not isolation.
They ask builders running agents with real money: where does your budget actually live, and have you tested whether the agent can get to it?
More from coding & agent
- Harness Self-Evolution Needs to Retro Its Own Retro Changes — mattpocockuk · 2026-11-03
- Open-source AI Xianyu monitor with Playwright + LLMs hits 16k GitHub stars — Usagi-org · 2026-11-03
- Researchers Wrote a Paper Assessing Matt Pocock's /retro Skill for Coding Agents — mattpocockuk · 2026-11-03
- Developer kicks off a Worms remake built with AI coding tools — ilumine_ai · 2026-11-03
- Creator hires an AI employee 'Viktor' in Slack to fix Claude's teamwork gap — rubenhassid · 2026-11-03
- Surrogate Inference Engine Accused of License-Washing Open-Source Project NInfer — teortaxesTex · 2026-11-03