The bottleneck moved from reasoning to execution

  • TS Rust's model usage represented an estimated $500–$800 of subscriptions, but agents also needed machines to build and test the code. Theo recruited stored computers and installed T3 OS on them. 22:28 23:00
  • The setup reached about eight local computers plus two VPSs accessible by SSH. Theo excludes those machines and their electricity from his later compute-cost figure. 23:26
  • Under contention, a ten-minute job could wait up to an hour before starting. More agent activity could therefore increase queues without delivering prompt validation. 24:14

What Blacksmith changed

  • Theo used Blacksmith to offload jobs to cloud compute. He discloses that Blacksmith is a sponsor and says its intended strengths are predefined repository jobs such as GitHub Actions. 23:26 23:48
  • He spent over $2,000 on TS Rust's offloaded test compute. This was separate from ordinary CI, which he estimates cost under $30. 24:14
  • Reported waiting for compute per run fell from upwards of 12 hours to 1.2 hours once agents could create and run cloud jobs. This is a waiting-time comparison, not a claim that the compiler itself became ten times faster. 25:05
  • Builds still took 20–40 minutes before their results could be checked. This remaining cost motivated Lust, a compiler focused on speeding up TS Rust's development loop. 25:30

Boundaries of the recommendation

  • Theo recommends Blacksmith for established verification, CI, and release jobs. He specifically discourages using it for arbitrary local agent development loops, except to check existing CI work. 25:54
  • The next approach, Boat sandboxes, addresses agents that need disposable machines on demand. 26:49 27:12