AI Platform

Agent Harness: private AI on your own machine

In developmentEngineered and documented, not yet released for sale. Figures are design targets.

Run open-weight AI models and agent workflows on hardware you control. Pick a model that fits the machine, describe the goal, watch a bounded team of agents work through a visual workflow, and approve the result before it leaves. No compulsory subscription to a hosted model.

  1. 01Your machineA laptop to start, a single-GPU workstation to grow, several machines later.
  2. 02Local modelsOpen-weight models served locally through Ollama, tested for instruction following, structured output and tool use.
  3. 03Bounded agentsMany agent roles, a set number of models running at once. The harness never promises more than the hardware can run.
  4. 04Visual workflowsEach step shown as it runs, with every event journaled for replay and audit.
  5. 05Human sign-offNothing is exported until a person approves it.

Private by default

Strict mode blocks every hosted model, so documents and prompts stay on the machine.

Honest about hardware

Models are matched to the memory you have and their measured speed is shown, not guessed.

Built for engineering work

Precifarm uses it to turn charging-site briefs into calculated, checked engineering packages that a registered engineer then reviews.

Specification

Agent Harness specification
ItemValue
ModelsOpen-weight models run locally through OllamaDesign target
AgentsA bounded team of agents with a set number of models running at onceDesign target
PrivacyStrict mode blocks every hosted model; you approve results before exportDesign target
HardwareStarts with small models on an ordinary laptop; larger models need more memory or a GPUDesign target

Figures come from Precifarm's engineering register (updated 2026-11-01). Design targets are not test results, and no certification is claimed unless a row says Tested.

Where it fits

  • Engineers and consultancies who cannot send client documents to a hosted AI service.
  • Startups that want agent workflows without per-seat model subscriptions.
  • Companies that need an audit trail of what an AI system did and who approved it.

What it is and is not

  • In development. Early access is by request.
  • Precifarm Agent on this website is a separate, hosted assistant and does not run privately.
  • Results are drafts until a qualified person reviews them; the harness never approves its own work.

Want private AI agents for your team?

Tell us about your hardware and the work you want to automate.