No Token Budget: Agentic Software Factories on Kubernetes

Ross says models had reached a capability threshold for handling the software lifecycle, but company development had not yet accelerated as expected. He frames the gap as a problem of implementation, architecture, and orchestration, and asks how far agents could go if given a greenfield project. The project tested whether agents could build and operate software continuously, with no token budget constraint.

The architecture models agents and deterministic processes as actors with unique identities, behaviors, and inboxes. Actors communicate through three actions: tell another actor something, spawn an actor, or listen for work to finish. A control plane places actors into suitable runtimes, while heavier development agents receive hermetic environments with tools such as Bazel, Docker, Kubernetes, and Playwright.

Ross describes single problem optimization, where workers generate candidates and a grader compares them and selects improvements. Candidates were complete hermetically deployed application instances, giving graders visibility into system behavior. Long-running objectives included reducing production errors, improving code quality, and runtime performance; finite objectives included features that could be completed and then retired.

Agents operated across isolated CI, pre-staging, staging, and production clusters, with each agent working against its own runtime. The agents also monitored system health and had authority to remediate issues such as Kubernetes restarts, database fixes, and network errors. Production error spikes could feed back into optimization cycles to prompt longer-term code changes.

Ross argues that teams should give agents the tools and isolated environments they need, rather than adding elaborate constraints. He emphasizes hermetic testing, canonical technology stacks, and ongoing background cleanup to manage codebase entropy. He says investments in infrastructure, developer experience, testing, and releases remain central to improving agent performance.