Agent systems can inspect models, tools, retrievals, costs, and outcomes. When a person enters the system, that shared operational picture becomes incomplete.

The person is usually represented by an action: approved, rejected, corrected, intervened. The changing human process that produced the action remains outside the system.

Human Runtime, or HRT, closes that gap.

HRT converts authorized human-performance inputs into interoperable states, spans, events, and traces. It gives agents and operational systems a structured way to work with relevant changes in human capacity, perception, behaviour, and action.

Observability is the starting point. It is not the limit.

Designed delegation defines who may decide. Human Runtime shows whether the decision point worked. It provides an observability substrate that policy, workflow, interface, and evaluation systems can use without collapsing measured capability into authority.

From observable to integrated

A system must perceive a relevant change before it can respond intelligently. Once human-runtime information is available, it can support more than monitoring.

  • An agent can calibrate confidence in a human intervention.
  • An interface can adapt to available attention or context headroom.
  • A training system can compare performance across attempts.
  • A safety system can identify a degrading handover before the outcome is known.
  • An operational workflow can route a decision to a person who currently has the right capacity and context.
  • A product can create new services around human–agent coordination.

The same HRT output can create different value in different environments. A measured reduction in context headroom may affect a pilot, an AI reviewer, a technician, or a racing driver in very different ways.

HRT supplies the common structure. Each vertical supplies its domain knowledge, operating constraints, and product logic.

That division matters. Authority is assigned through policy, governance, law, and organizational mandate. Capability is observed under changing conditions. HRT can inform an authorized routing or intervention policy; it must not silently decide who is entitled to act.

The HRT stack

The stack connects inputs from the human and the task environment to systems that can act on them.

  1. Inputs — sensors, interfaces, behaviour, task events, and existing software telemetry.
  2. Core — the open protocol, processor, shared data model, confidence model, and reference services.
  3. Integration — APIs, events, agent tools, policies, and bridges into existing telemetry.
  4. Verticals — domain-specific interpretation, workflows, integrations, and products.

This structure is deliberately horizontal. The core does not need to know how to fly an aircraft, supervise an AI agent, run a factory, or coach a driver. It needs to represent what can be known about the human runtime—with timestamps, definitions, provenance, uncertainty, and validity conditions.

Companies can then integrate HRT into existing systems, combine its output with proprietary data, or build vertical products on top of the stack.

One canonical vocabulary

The public concepts and machine identifiers reinforce each other.

Human Node describes the person as an active, changing part of an intelligent system. Inside the protocol, that node is identified as hrt.node.

The canonical model is deliberately small:

  • Human Node — hrt.node: a person as an active, changing part of the system.
  • Runtime State — hrt.state: a time-bound estimate with confidence, provenance, freshness, and validity conditions.
  • Runtime Span — hrt.span: a bounded period of human activity or interaction.
  • Runtime Event — hrt.event: an observed action, change, intervention, or handover.
  • Human Trace — hrt.trace: connected human, agent, tool, environment, and outcome events on a shared timeline.

These objects do not claim access to private thought or complete knowledge of cognition. They represent measurable, operationally relevant observations and estimates. Uncertainty is part of the object rather than an inconvenient footnote. More specific profiles—cognitive, perceptual, physiological, or behavioural—can extend the model without redefining the core.

The result is a vocabulary agents can query, systems can exchange, and people can audit.

Extend the systems already in use

HRT connects to the agent and observability ecosystem rather than asking teams to replace it.

  • Runtime spans and events can map into OpenTelemetry conventions while preserving HRT-specific meaning, confidence, consent, and provenance.
  • A provider-neutral collector can receive measurements, apply policy, and export authorized HRT data through existing telemetry pipelines.
  • Agent protocols can advertise or request HRT context through bounded extensions instead of silently assuming a person's state.
  • Operational platforms such as Datadog can receive correlated HRT telemetry alongside model, tool, service, and machine traces.

The canonical HRT record remains the source model. Each bridge carries the relevant parts into an external system without letting that external system define HRT.

Vertical products build above the protocol

HRT is not a single aviation product, human-in-the-loop dashboard, sensor, or cognitive analytics package.

Those are possible implementations.

An aviation company could use HRT to improve handovers, training, or crew–automation coordination. An agent platform could use it to interpret approvals and escalation. A racing organization could connect driver state to vehicle telemetry. An industrial operator could combine HRT with alarms, procedures, and control-room events. A training company could build adaptive instruction around the same protocol.

Some partners may expose HRT directly. Others may integrate it invisibly into a larger product. Both expand the value and reach of the stack.

This creates an ecosystem rather than a catalogue of isolated use cases:

HRT provides the shared human-runtime layer. Vertical builders turn that layer into domain value.

The reference point for Human Runtime

humanruntime.org is the reference and coordination point for this emerging layer. HRT is currently an independent, founder-led research and engineering initiative. The direction is an open protocol, reference implementations, evidence, and governance that can become increasingly community-led as the work earns adoption.

It brings together the core definitions, protocol, processor, research, failure analysis, integration patterns, and vertical applications. It can document what HRT means while allowing many companies to create products from it.

The Topic Sorter supports that work. It helps investors, partners, and adopters learn the concepts, relate them to a specific environment, and reveal promising integrations. It is an onboarding and discovery tool—not the HRT core.

Developers and scientists can test a schema, processor, integration, or measurement source without joining a closed vendor ecosystem. Companies can build proprietary vertical value without owning the shared language. Investors and institutions can fund reusable infrastructure and domain pilots in parallel.

The larger objective is straightforward:

Make the human runtime legible, interoperable, and useful inside agentic systems.