@ClaudeCodeLog The interesting part is the shift from “agent as a prompt” to “agent as a configured worker.”
Dirs, permissions, model, effort, plugins/MCP are the contract surface. Once explicit, teams can reason about what an agent may touch, how hard it should think, and how it recovers.
@ClaudeDevs The underrated part is not just lower latency.
For agent workflows, cold start becomes a product cost. Every pause before first token makes the loop feel less delegated, more like chat.
Prompt cache helps agents feel like a standing work surface, not a fresh conversation.
@OpenAI The important part is not “coding from your phone.”
Mobile turns Codex into a remote work control surface: start work, check state, approve next steps, and keep the loop alive while the compute stays elsewhere.
That is a more useful shape for agents than another chat box.
@claudeai The interesting bit is not tiny computers as swag.
It is giving builders a constrained physical work surface: state, sensors, feedback, and failure modes.
That is the layer agents need before they feel like operators, not just chat models that produce plans.
@OpenAI The real reason to bring Codex into a team is not “more AI code.”
It gives agents a managed work surface: repo context, tasks, reviews, approvals, and verification.
That is where coding agents move from clever autocomplete to accountable engineering workflow.
@ClaudeDevs The interesting part of /goal is not persistence.
It turns chat into a work contract: keep context, make progress, survive interruptions, and report back when the loop changes.
Real agent work is mostly waiting, checking, recovering, and closing the loop.
@bcherny The underrated part is not just Opus 4.7 being smarter.
Flight booking is a real work surface: preferences, constraints, vendor UI, timing, confirmation, and edge cases.
Once agents can close that loop reliably, they stop feeling like copilots and start becoming operators.