I’ve been working on a developer-facing sandbox for AI agents that focuses on budgeting and cost control, not payments.

In multi-agent setups, we kept running into issues where agents either hoarded resources or exhausted shared budgets unpredictably. So we built a control layer where agents operate using virtual credits, can temporarily rebalance budgets or split shared API costs, but everything stays under explicit human-defined limits with full audit logs and kill switches.

It’s intentionally not real money and not a financial product — more like infrastructure for coordinating agent spend safely. Mostly exploring how much autonomy you can give agents before cost becomes the real bottleneck.