Tag
An opinion piece arguing that AI agents should never hold raw provider API keys, but instead be given prepaid, capped credentials with per-hour, per-day, per-tool limits, tool allow-lists, price checks, and per-call receipts. The author uses the Agentik setup as a reference and asks readers whether they prefer raw keys with monitoring, hard limits, or something in between.
A discussion about whether AI agents should be given their own spending budgets, such as a monthly cap to call paid specialist APIs, versus hardcoding vendor access. Asks practitioners for real-world experiences and failure modes.
A compromised cloud account resulted in an $80,000 AI bill, underscoring the critical gap in AI providers' ability to enforce hard spending caps and protect against both malicious hacks and uncontrolled AI usage.
The article describes a capability layer for AI agents that lets them discover, pay for, and verify external services with strict controls like spending limits, durable receipts, and replay protection, applied to EU compliance tools.
Discusses the idea of giving AI agents a small, locked-down budget (e.g., $200) for routine business expenses like software trials, drawing a parallel to giving a junior employee a limited corporate card.
datasette-llm-limits 0.1a0 is a new plugin for Datasette that allows configuring per-user or global spending limits for LLM usage, working alongside datasette-llm and datasette-llm-accountant.