Give the Agent a Budget, Not a Token — Anthropic's Sachin Malhotra
AI Engineer · youtube · 2026-08-22
Anthropic engineer Sachin Malhotra shares a production incident where an agent deleted 200 workloads in 90 seconds. The root cause was granting unbounded token-based permissions. He argues for replacing boolean tokens with a 4-dimensional Budget: how much, how fast (rate limits), what can be undone, and who notices.
Key practices include:
- Asymmetric Verbs: Grant autonomy on operations that fail loudly (e.g., unskip tests), keep humans on silent failures (e.g., skip tests).
- Rate Limits: Self-refilling ceilings on writes to avoid friction.
- Trip Wires: Use aggregate counts to trigger rules instead of static allow-lists.
- Undo Test: The ultimate metric for safety.
- Identity Stamping: Enforce identity via a proxy to prevent agents from resetting limits by spoofing headers.
More from coding & agent
- Perplexity open-sources SPACE, the production sandbox platform behind Perplexity Computer — AravSrinivas · 2026-10-16
- Meta releases Muse Spark Open: 816B MoE, 42B active, 1M context, day-0 on Baseten — baseten · 2026-10-16
- Perplexity open-sources SPACE, the sandbox platform powering its Computer agents — AravSrinivas · 2026-10-16
- Vercel makes Python generally available: FastAPI, Django, Fluid Compute support — steren · 2026-10-16
- Full course: build a playable Unity fishing game with Claude Opus 5.5 in six steps — Roger_M_Taylor · 2026-10-16
- Marc Lou praises PostShiba, a cute indie alternative to AWS SES for email infra — marclou · 2026-10-16