GitHub gated cloud agents. I'm fixing the interaction layer.

GitHub added per-organization controls for Copilot cloud agent in the last 24 hours. OpenAI also updated its Agents SDK for safer and more capable enterprise builds. That is where the market is moving now. More agents. More surfaces. More control. Less patience for rough edges. That part matters to me because the work across the last day kept running into the same problem from every direction. Desktop and mobile prototype parity. Mention pills that still fight backspace. Text inputs that need to stay in the user zone and stop trying to be clever. Right-click behavior that still decides whether the terminal feels real. Git pane structure that still needs calmer logic. Thread history that still does not carry forward cleanly. So this stream is about the interaction layer. Not a redesign for the sake of it. The layer that decides whether the product feels trustworthy once the model is already there. If the mention flow is awkward, the tool still feels fake. If the text input keeps nudging the user instead of getting out of the way, trust drops. If the session history does not survive the jump, the whole agent story breaks again. The receipts were loud today. 134 merged events across Letta, Claude Code, and Codex in the last 24 hours. Parallel work across desktop, mobile, git, terminal, and the surrounding rules. The work kept expanding. The interaction quality still needs to catch up. I am back in the room to make those edges feel right. Less noise. More usable defaults. Sharper parity between prototype and product. If you code, you probably know this kind of session. The model is not always the blocker. The layer around it is. 8 to 10 hours. AI radio on. Quiet co-working. Just the real interaction work. GitHub: https://github.blog/changelog/2026-04-15-enable-copilot-cloud-agent-via-custom-properties OpenAI: https://techcrunch.com/2026/04/15/openai-updates-its-agents-sdk-to-help-enterprises-build-safer-more-capable-agents/