Charles: agent behavior is hard to control, so permissions define this abstraction
charles_irl · x · 2026-09-10
The core tweet of Charles' thread: agent behavior is much less controlled — intelligence is powerful when given general tools and broad access, but general tools can surprise you — so this incarnation of the abstraction will revolve around permissions.
More from coding & agent
- bury-bench: a deterministic, zero-LLM-judge benchmark scoring coding agents on ADHD-friendly answers — Obluness · 2026-09-11
- Gemini API Docs Move Into AI Studio, with Docs MCP and Agent Skills Support — _philschmid · 2026-09-11
- ZenML built its first RL environment to benchmark coding agents on ML workflow tasks — strickvl · 2026-09-11
- Reviewer: Relume is the first AI-first design tool that doesn't skip the design process — michalmalewicz · 2026-09-11
- OpenAI agents hijacked a German wiki to share notes — dev builds Pod, an MCP for agent tool reviews — saheedniyi_02 · 2026-09-11
- Greylock: the next AI security problem is agent intent — SethGRosenberg · 2026-09-11