What is Harness? | HarnessHarness owns the agent loop between a model API and an explicitly bounded set of tools.Harnessreferenceharnessreferenceadopterdeveloperoperatorrun-agents

One loop, explicit effects

ReferenceHarness

ForAdopter · Developer · Operator

Source & revision

harness/website/docs/index.md · 90f10a4314c1

Harness is the beyond10x agent loop. It talks to model APIs directly, assembles each turn, executes tool round trips, asks for approvals, enforces budgets, and records what happened.

It is deliberately not a wrapper around a vendor's coding-agent binary. If you need to drive and compare vendor harnesses, that is metaharness. Harness owns the loop itself.

user or application
        │
        ▼
  b10x-harness ───────▶ model API
        ▲                   │
        │                   │ tool calls
        └── approval gate ◀─┘
                 │
                 ▼
       workspace tool catalogue

What it owns

Harness keeps five concerns together because they all decide what one run means:

• Turns. A stateless conversation is projected onto either the OpenAI Responses or Anthropic Messages API.
• Tools. The model sees exactly the catalogue this machine can perform.
• Approvals. Calls above the run's unattended risk ceiling are decided before the effect.
• Budgets. Turn, token, duration, context-window and—when rates are supplied—cost ceilings bind the loop.
• Records. Sessions preserve a conversation locally; JSONL events expose the live run to another program.

The capability ladder

The command line starts read-only. More consequential tools appear only when a confinement boundary can provide them.

Machine boundary	Published catalogue	What the run can do
No substrate	file_read, dir_list, search, find	Inspect one workspace
Confined workspace	the read tools plus file_write, file_edit	Change admitted paths
Confined execution	all of the above plus run	Start an explicitly allowed argv

A missing capability is not hidden. b10x-harness tools and the first JSON event report a tool that was requested but withheld, together with the machine fact that withheld it.

Three ways to use the same loop

Surface	Use it when
b10x-harness run, chat and workflow run	A person, shell script or evaluator owns the process
harness-loop library	Another Rust component should bind model and tool ports in-process
b10x-harness app-server	A compatible client owns tools and talks JSON-RPC over stdio

The library core has no terminal policy. Its default approver is deny-all; each shell must choose its own interaction policy explicitly.

Where to go next

Start with Getting started for a read-only first run. Then read The agent loop to understand the execution model or Tools and approvals before enabling effects.

:::info Pre-v1

Harness is tagged 0.12.1 and changing quickly. The provider wires and command-line surface are pinned by repository contracts, but live-provider and external-bridge characterization is still limited. See Status and limitations before adopting it.

The source is publicly readable under LicenseRef-B10x-Proprietary. Public visibility is not an open-source licence or a stability promise. See the repository security policy for private vulnerability reporting.

:::