Press adds first-class Effects and adversarial approval for agent tool calls
RichmanRonald · x · 2026-07-23
The next version of Press turns “Effects” into a first-class mechanism for external, irreversible state changes in agent workflows.
It adds an adversarial session by default to run all external effects, with optional policies written in the same Markdown/OpenProse structure as the rest of the system.
The post frames this as a direct answer to safety problems such as model-to-tool misuse: instead of trusting the main agent, a separate model can sit between tool calls, review the trajectory, and approve or block actions.
More from coding & agent
- Codex stress tests on old repos reveal a new idea: an agent for the margins — joonasvirtanen · 2026-07-23
- The Turing Post says most AI agents still do not need graphs — TheTuringPost · 2026-07-23
- Repomix packs a codebase into one AI-ready file with filtering, compression, and token counts — KhuyenTran16 · 2026-07-23
- Motion UI lands in the shadcn registry with CLI, prompt generator, and MCP support — shadcn · 2026-07-23
- Open Teach aims to teach any agent from browser activity and custom data — techNmak · 2026-07-23
- Git may become obsolete if AI agents can coordinate through a new communication layer — triviatroy · 2026-07-23