Blog

A phone can do more than monitor Codex. Hermes shows a second pattern: run your own agent on a VPS, connect a GitHub knowledge base, and use Telegram or Discord—with a separate identity, isolation, and audit.

MIT EQuS connected Codex with GPT-5.6 Sol to live laboratory hardware. The agent completed 36 of 40 target measurements without intervention and monitored a 12-hour overnight run—but noisy signals show why physical systems need stop conditions and human judgement.

A long AGENTS.md is not automatically better. Reliable agents need concise rules, current sources, permission to acknowledge uncertainty, and mandatory result verification.

Grep finds text. LSP gives a coding agent definitions, references, and symbol relationships. That reduces blind tool calls, token use, and false edits during refactors—especially in larger codebases.

OpenAI is expanding Codex with Sites, role-specific plugins, and team workflows. This is less about generating a page and more about creating useful internal work artifacts.