Before you press play
What you’ll find in the video
- Break a workflow into small questions that code can combine, such as which messages need an agent or when it should speak; a fixed output shape does not make the judgment deterministic or correct.
- Josh describes putting decision models inside the harness—the software that runs the agent—and exposing them to plugins, rather than relying only on an LLM to call Jev as a tool. These integrations are exploratory, and coding agents still struggle to write clear atomic questions.
- Kevin separates trusted configuration and control from untrusted agent execution, with sandboxing and agent identities as additional boundaries. Enterprise deployment needs those controls, not just a prompt asking an agent to behave.
Worth knowing
Recorded September 30 and published October 1 UTC. This is a discussion, not a tested installation guide. At recording, harness features were experimental and OpenClaw Enterprise was pre-1.0, intended for internal pilots rather than established production readiness. DOOM illustrates composing decisions in a real-time loop, not the best way to build a game bot; speed, cost and security claims were not independently tested.