Engineering blog
Engineering notes
Implementation-level writing on disposable compute, agent execution, batch systems, and reinforcement learning infrastructure.
Agent infrastructure
9 min read
A coding agent needs a computer, not another tool wrapper.
Build a disposable workspace for OpenCode, Claude Code, Codex, Deep Agents, or Aider, then make its lifecycle explicit.
Agent infrastructure
8 min read
Disconnect the terminal. Keep the agent running.
Launch a durable process inside a sandbox, recover its logs and status later, and clean up the machine deliberately.
Batch systems
10 min read
Treat every CI shard as a disposable machine.
Fan out isolated test jobs with bounded concurrency, preserve command failures, and destroy every sandbox on every path.
RL systems
14 min read
One sandbox can run RL. A pool is built for the loop.
Compare local vector environments, a complete training job in one sandbox, and one batched request driving N warm environments.