Johannes Vermeer, A Lady Writing, c. 1665. National Gallery of Art, 1962.10.1. CC0.
The company that runs on agents.
Good morning.
We missed some weeks. That is on us. No soft landing: the Friday letter did not ship, and the desk went quiet while the stack fell over.
Keeping connectivity alive without a human in the loop is still an unsolved problem. When AI tools lose auth mid-series, the whole digest chain stalls. It is disruptive and annoying. Call it what it is — an ops failure of the stack, not a mysterious creative pause.
And while we are being honest: we know each week is mostly AI slop. The point of this letter is still a better digest of the bookmarks and a short list of things worth following up on. Synthesis over a raw dump. That is the bargain. We are back to keep it.
September's bookmarks argued one thesis, over and over: stop treating agents as solo laptop sessions. Wire a factory. Cloud sessions anyone can join. A Chief-of-Staff agent that surfaces review queues. MCP as the connector layer. An immutable box the agents cannot trash. Non-eng teams running the same ticket-and-agent rituals as eng. That is the company shape under the backlog — not another chat product.
Factory harness, not solo chat.
Inside OpenAI, Codex/Work took over eng and non-eng. Human sets the outcome. The agent gathers repo, Slack, Notion, Datadog. Implement, CI babysit, multi-lens agent review, watch the rollout. Perf Factory, deploy handholds, Sevbot. External Codex is weaker than a fully wired internal harness. Steal the loop shape. Do not copy the unlimited-token budget.
Sergey's team runs Claude, Codex, and Pi in the cloud for a year. Any session is joinable. Meetings spawn research and implement subagents. A Chief-of-Staff agent owns the review queue and unblocks people and agents. Multiplayer cloud as default — not a lonely laptop tab.
Rhys on why MCP beat CLIs: indexable tool catalogs, consistent auth, multi-account, code-mode return shapes. Late takeoff was bad clients, not a bad protocol. Treat connectors as a product surface, not a pile of shell scripts.
The whole company, on the same rituals.
Warp's COO: marketing, benefits, people, talent, partnerships, ops — none are eng — all run on Linear plus Claude Code with eng rituals. Tickets, agents, and the same operating cadence. Steal the transfer. Do not wait for a "non-eng AI product."
AgenticLinux: immutable Fedora bootc desktop, agents and Docker preinstalled, atomic Hub updates with rollback. Agents mutate sandboxes, not /usr. A dedicated always-on box without lid-close drama.
Southbridge wired TypeSafe's Jev into an entity-resolution pipeline and cut model cost ~99% versus a frontier baseline, staying near Fable accuracy with much higher throughput. Use a cheap system-one workhorse inside a staged hank. Tight evidence criteria. Right-sized models for messy joins.
Try this
Three moves. Build the factory, not another chat.
- Pick one real workstream. Run it as a cloud session anyone on the desk can join. Put a CoS-style agent on the review queue — not another solo tab.
- Inventory connectors. Prefer MCP where the catalog and auth stay consistent. Kill one-off CLI wrappers that only one person understands.
- For the next messy join or normalize job, stage a cheap system-one model first. Refer hard cases up. Measure cost and accuracy before you default to frontier.
Back next Friday — with fresh keepers, if the stack holds.