Codex went enterprise. I pitched what agents are.

OpenAI announced Codex is scaling to enterprise through Accenture, Capgemini, TCS, and five other global integrators. ChatGPT Images 2.0 also shipped. Augment published their intent-based pricing model. The enterprise agent tooling market is consolidating fast. Yesterday I gave a talk at Claude Istanbul. 23 slides. No product pitch, no demo reel. Just an honest explanation of what a coding agent actually is. The thesis: a coding agent is a runtime assembly, not a fine-tuned brain. Training gives you the base. Fine-tuning adjusts it. But the thing that makes an agent useful, the system prompt, the memory, the tools, the skills, the reference repos, that's all assembled at runtime. Every session is a fresh construction. The model never changed. The context window did. That's what tokenmaxxxing means. Not cramming more tokens into a prompt. Understanding that the quality of what fills your context window is the entire difference between a useful agent and a confused one. Before the pitch, 234 events across 20 sessions. The real work was a full ACP sprint. Terminal splits are wired. The composer squeezes into a footer when a TUI takes over. The setup assistant pane is live. Mode manifests got a real design. Gemini model activation landed in the desktop. Five reference repos got studied for frontend state management. The prototype parity plan is written. The hardest unsolved problem right now: keyboard shortcuts get swallowed when a browser pane is open. Command palette, devtools, all blocked. That needs a proper focus routing fix, not a workaround. Tonight is a shorter session, two hours. The priority is ACP wiring. The desktop needs to become the browser for agents the same way Chrome became the browser for websites. That means every agent, regardless of which runtime it came from, should be inspectable, composable, and interactive through the same surface. If you write code and the pace of all this makes your head spin, same. The way through is the same as always: build the next thing, verify it works, keep going. AI radio in the background. Come build. Sources: Codex enterprise scaling: https://openai.com/index/scaling-codex-to-enterprises-worldwide/ ChatGPT Images 2.0: https://openai.com/index/introducing-chatgpt-images-2-0/ Augment intent pricing: https://www.augmentcode.com/guides/intent-pricing Today's pitch: https://compounddd.com/presentations/graphyn-claude-istanbul-2026-04-21-xray