Module 24 · 60 min

Ship It: Hosting, Cost, Observability

You can size a host from the subprocess model, cap a session's spend with real limits, and say which cost numbers are safe to bill from and which are not.

Workbench tag
module-24
A thick wall with five slots set into it; a pale card is emerging from three of them while two are shuttered closed.
Draft

This page is not written yet. The outline below is the plan of record from curriculum/BUILD_TRACK_PLAN.md; the prose, the code, and the verified block land with the workbench’s module-24 tag.

What this page will cover:

  • The subprocess model as your sizing driver: one session, one claude process, one shell and working directory. Fifty concurrent sessions are fifty processes, yes, really.
  • The hosting page’s four session patterns by their own names: ephemeral, long-running, hybrid, multi-agent container, and pinning long-running sessions to containers by hashing the session id.
  • The resource floor that page now states: 1 GiB RAM, 5 GiB disk, 1 CPU per agent as a starting point, sized up by measuring peak memory on representative sessions. Token cost, per the same page, usually exceeds container cost by an order of magnitude.
  • maxTurns and maxBudgetUsd as operational limits, the /clear reset, the asymmetry in what the budget-exceeded error counts, and the cost-tracking page’s warning that these fields are insight, billing belongs to the Usage and Cost API.
A sealed pipe with a small tap fitted to its side, drawing a thin sample line off into a tray while the main flow continues past untouched.
  • OpenTelemetry wiring, and the build-versus-buy fork the hosting page itself names: Managed Agents, where Anthropic runs the agent and its sandbox.