Agent Harness
Agent Harness is an execution layer for AI coding workflows.
It is built around one simple operating model:
- use a strong supervisor model for planning and review
- delegate bounded tasks to cheaper or local workers
- verify results deterministically
- keep the operator in control when approval, failure, or recovery matters
What it is not
It is not trying to be a general chat UI, a hosted control plane, or a vague “agent platform.”
The current product path is narrower and more useful:
- standalone local-first operation through the CLI, SDK, MCP, or local UI
- reliable local and mixed-model execution
- evidence-rich artifacts
- benchmarkable routing and retrieval improvements
Public documentation boundary
These docs describe shipped, supportable behavior: installation, CLI commands, configuration concepts, MCP and loopback integration, operations, and recovery. They intentionally omit private prompts, credentials, internal deployment addresses, unreleased experiments, and implementation details that are not part of the public contract.
Product priorities
The maintained roadmap focuses on standalone reliability, evidence-rich execution, benchmarked routing and retrieval, and a possible future path toward shared operation. See the roadmap for the current direction; the numbered phases in internal planning documents are not a public release schedule.
◆Current best-fit user
The first market is solo developer power users running local models and wanting a trustworthy harness around them.