Skip to main content
The Voss harness is a terminal-native agent loop for bounded, inspectable, resumable repo work. It is the single-agent substrate the orchestration layer builds on: a team run is many of these loops, each a scoped node in a session tree. It is not a general chat box. Each turn goes through a small set of explicit systems:
The confidence check reads the plan’s own reported confidence. Below the threshold (0.60 by default), Voss asks you a clarifying question instead of executing. That number is the model’s self-report and is recorded as such; it is not a measured probability.

​
Why it exists

AI coding tools are useful, but many hide important decisions in prompts:
  • What can the agent edit?
  • Which commands can it run?
  • What context did it inspect?
  • Which files did it avoid?
  • Can the session be resumed later?
Voss makes those decisions explicit through commands, modes, tool classification, permission prompts, persistent project context, and resumable sessions.

​
Main commands

voss do

Run one natural-language task and exit.

voss

Open the full-screen Go terminal client, or voss chat for the plain REPL.

voss edit

Open a scoped edit session around explicit targets.

voss sessions

List and resume prior harness work.

voss serve

Expose the same loop over a local REST + SSE protocol for SDKs, swarms, and Voss ADE.

​
Product shape

The harness is the substrate; the orchestration layer is the product built on top of it. One harness loop is a single bounded agent. A team run dispatches many of them under an Engineering Manager loop, each owning a scoped, budgeted node in the session tree, gated by independent review. The Python harness is the canonical behavior source and the npm CLI vendors that implementation. Clients sit on top of it: the Go terminal client and the Rust, TypeScript, and Go SDKs all talk to voss serve, and so does the separate Voss ADE desktop app. None of them reimplement harness behavior. The .voss language is the control layer at both levels — confidence, context budgets, routing, memory, tools, and fallbacks for one agent; roles, principles, and gates for a team.