harnessie

Harnessie is an open-source framework designed for verifiable orchestration of multi-agent AI workflows, emphasizing user control and auditability. It addresses the risks of unsupervised AI by ensuring that every action taken by an AI model is subject to independent verification and explicit user consent before any side effect occurs, such as changing a file or running a command. No sensitive data leaves the models controlled by the user, and every action—whether by AI or human operator—is recorded in a hash-chained audit log, making tampering detectable.

The framework structures jobs using three types of agents: an orchestrator that decomposes tasks, workers that execute tasks in isolated environments, and verifiers that independently check results. Each step in a workflow is separated by checkpoints, and the system defaults to failing closed, meaning no action is completed without explicit approval. When a judgment call is required, the process halts for human intervention, ensuring the final decision remains with the operator. Disagreements between agents are recorded rather than averaged, and declined actions are logged without being overridden.

Harnessie is brain-agnostic, allowing users to switch between local, private, or remote AI models—including those accessible via OpenAI-compatible endpoints—by editing a single configuration file. The framework supports eight providers and eleven verified models, and its structure remains consistent regardless of the underlying model. Installation requires Python 3.11+ and PyYAML, and can be performed using pip, pipx, uv, or brew. Model adapters rely on the standard library, with no dependency on vendor-specific SDKs.

The tool is self-hostable and released under the Apache-2.0 license. It includes a test suite and evaluation scorecard that can be run against a deterministic mock environment without network access, allowing users to validate its operation before integrating any API keys. Harnessie is suited for developers and operators who need strict control, verifiability, and auditability in multi-agent AI workflows, particularly where sensitive data and independent verification are priorities.