A new category for governed human-AI collaboration. Humans, AI, and hybrid actors executing against a single primitive — with one audit trail for all of them.
Models are rented. The harness is owned.
Most AI products embed operational intelligence inside prompts, runtimes, or orchestration code. All three are ephemeral. Models deprecate. Prompts drift. Orchestration logic lives only inside the particular agent framework that wrote it.
Inistate makes a different commitment. We externalize workflows, states, forms, transitions, and audit trails into persistent schemas. Those schemas become the operational source of truth . Humans, AI, and hybrid actors all execute against them, through one primitive, with one audit shape.
We call this substrate the harness. It is the category Inistate defined and the category we lead.
The schema is what the AI reads to know what is possible. The schema is what the AI writes to record what happened.
Every operation against the harness is a
State → Activity(Form) → StateEvery transition produces one history event with identical fields: by, on, changes, and an optional ai object. The audit shape does not vary by actor type. Compliance reporting is a query, not a reconstruction.
Workspaces, modules, schemas, transitions, and tools are all discoverable through the MCP server. The AI does not need pre-loaded context to operate inside an unfamiliar deployment. It bootstraps through the protocol.
AI-submitted transitions persist confidence scores. Below threshold, the system records an intention event with flagged: true and the AI must escalate to a human. Regulation-grade human-in-the-loop, built in.
Activities are typed by who can perform them: human-only, AI-only, hybrid, or any-actor. Promoting a human activity to AI changes one field in the configuration. The process model, the form schema, and the audit are unchanged.
Each of these is shipped. Each is documented in the MCP Server Specification v1.0. Each is exercised in the seven proofs.
The six properties are not a checklist. They are a dependency graph. Each layer requires the ones beneath it. Pull any layer and the layers above it collapse.
A competitor who wants to reach parity cannot copy one feature at a time. The features depend on each other. To reach parity, a competitor must rebuild the substrate in the correct order — an architectural commitment that breaks installed customers and consumes years of engineering capacity.
Vendors with adjacent products will tell you they ship a harness. They do not. Each of the four labels below describes a product the harness includes properties of, but none describe what the harness is.
A workflow tool.
Workflow tools have one actor type: humans through forms, or services through APIs. The harness has unified actor parity from the primitive up.
An agent framework.
Agent frameworks embed operational intelligence inside prompts and runtimes. The harness externalizes it into persistent schemas that survive the agent.
A low-code app builder.
Low-code app builders generate UI on top of stateful databases. The harness is a state machine with native AI execution and audit, with UI as one possible surface.
A model orchestration layer.
Model orchestration layers route prompts between models. The harness is model-portable; the workflow is the truth, and the model is swappable.
Seven demonstrations record the harness executing across orthogonal stress dimensions. Each holds different variables constant and varies others. Together they triangulate the architecture: it is not just correct in any one dimension, it is invariant across all of them.
User input"go to LV 00003 and Approve it"
The leave application LV 00003 was created by a human. Claude Opus 4.5 advanced it to Submitted. GPT-5.5 Thinking approved it. Each transition produced an identical history event.
Actor Parity across model families.
Models are interchangeable; the audit is not.
User input"build a leave application module"
In a single session, on Claude Haiku 4.5, the AI designed the schema, generated the form, configured the transitions, set up the audit, and ran a test entry through the full happy path.
All architectural properties composing simultaneously, on a small model.
Model size is not the bottleneck. Architecture is.
User input"track this"
A photograph of a blood pressure cuff. The AI parsed the image, inferred the schema, generated the module, and made the entry — without a text description.
The harness operates on artifacts, not just on language.
Schema-as-the-prompt generalizes beyond text.
The harness substrate is covered by US Patent 20230266946 A1 (issued August 2023). The Actor Parity governance extensions are covered by US Provisional Application #64/045,012 (filed April 21, 2026). A competitor must engineer around both layers to ship a competitive harness. The difficulty multiplies; it does not add.