Multiplayer AI still lacks a real permission architecture
edgarpavlovsky · x · 2026-07-28
We still don't have a good permission architecture for multiplayer AI.
The author argues that shared company resources need to be filtered through each user's permissions, whether the actor is a human or an agent. That means:
- The “company brain” should present different views to different users.
- The skills library should also be personalized by permissions.
- In some cases there are two permission layers: what the agent can access, and what humans are allowed to do with that agent.
The takeaway is that multiplayer AI introduces a new control-plane problem that software systems have not really faced before.
More from coding & agent
- IBM launches a free 1-hour graph engineering course for AI agents — goyalshaliniuk · 2026-07-28
- Opinion: Agents dynamically building specialized harnesses on the fly will be huge — burny_tech · 2026-07-28
- Agents will increasingly build task-specific harnesses for themselves on the fly — corbtt · 2026-07-28
- Swarms compares its GraphWorkflow with LangGraph on code size and parallel execution — KyeGomezB · 2026-07-28
- Custom SWE-bench on a Rails codebase finds Opus 5 tied Fable 5 on quality — sergeykarayev · 2026-07-28
- MCP’s next spec makes the core stateless and ships July 28 with Tasks and MCP Apps — burkov · 2026-07-28