Anthropic became AWS of agents. sync is the hard part.

Anthropic launched Claude Managed Agents yesterday. Hosted agent infrastructure with sandboxed execution, checkpointing, credential management, and multi-agent coordination. $0.08 per agent runtime hour. Notion, Rakuten, Sentry, Asana already on it. Cloudflare lost 13%, Fastly 18%, Akamai 13% in a single session. The market decided that Anthropic is now the deployment layer for agents, not just the intelligence layer. On the same day, OpenAI got hit by a supply chain breach. The Axios npm compromise from March exposed macOS signing certificates for ChatGPT Desktop, Codex, and Atlas. North Korean actors were behind it. If you ship desktop apps or use CI/CD pipelines, this is your threat model now. And OpenAI launched a $100/month Pro tier with 5x Codex usage, filling the gap between $20 Plus and $200 Pro. The coding agent pricing war is officially on. Last session was 289 merged events across Letta, Claude Code, and Codex. 12 separate Codex sessions running simultaneously. Six concurrent heavy sprint lanes across desktop, backyard, mobile, prototypes, and the design system. Most of the day was not hype work. It was multi-lane sprint execution live. The big thing from that session was sync pane design. Not patching a bug. Not shipping a feature. Designing the concept from scratch. Sync is the hardest unsolved problem in the app because it forces you to answer what the product actually is. What do we sync? If we are syncing machines, we list machines. If we are syncing coding harnesses, we list harnesses. The data model is people times machine, user root slash project level, harness slash agent. And on top of that, MCP stacks, skills, and lenses need to flow through the same pipe. The sync pane is not a settings page. It is where agent runtimes get configured. Where it is decided what runs where, with what capabilities, and how it all stays current across machines. tbh.md becomes the distribution layer for what gets synced. The architecture also crystallized this session: the user sits in front of a desktop app that talks to ACP, which talks to providers like OpenAI, which host agents like Codex, which run models. There is no backend in the middle. The agents kept trying to use hardcoded admin endpoints. The builder corrected them five times: prototype first, then build. This stream is about continuing the sync pane design and building the real data flow. If you build desktop apps and struggle with multi-machine state, if you have ever tried to sync development environments across machines and felt the concept collapse under its own weight, this is the same problem at a different scale. AI radio is on. If you are coding and trying to make state actually persist and sync instead of just look right in one place, this is the room. Sources: - https://www.anthropic.com/engineering/claude-code-security - https://www.anthropic.com/managed-agents - https://www.theverge.com/openai-supply-chain-breach-axios-macos - https://www.cloudflare.com/emdash