Sections
On this page
What a Harness is
A versioned, executable workflow in which external agents and people produce inspectable decisions and tracked Flownix work.
A Harness is not a chat or a prompt collection. It is a versioned process: roles, an executable step graph, policies, budgets, inputs, and outputs. The same published version can run repeatedly for different Flownix nodes.
- 01Version
- 02Run
- 03Work item
- 04Deliberation
- 05Decision
- 06Approval
- 07Flownix nodes
When to use it
Architecture and design
Independent roles propose options, test assumptions through claims and critiques, then vote or hand the choice to a judge.
Review and audit
A repeatable graph gives each agent explicit context, lease, retry behavior, output, and route to the next step.
Safe decomposition
An agent can return structured Flownix nodes, a person approves them at a human gate, and create_flow_nodes applies normal project checks.
Available today
The runtime validates and publishes immutable versions, runs agent and control nodes, routes by outcome, shares council deliberations and ballots, waits for human approvals, enforces tools policies and budgets, and exposes the complete run context through MCP. External agents do the model work; Flownix coordinates and records it.
The core principle
Inspect the process, not hidden reasoning. Proposals, claims, evidence, critiques, votes, decisions, step outputs, approvals, and created Flownix nodes remain available as explicit records.