Agent systems can trace model calls, tool use, retrieval, messages, cost, latency, and machine outcomes. When an agent asks a person to approve, correct, interpret, or recover a task, the operational picture usually collapses into a button click or text response.

That is not enough for dependable human–agent coordination.

HRT gives agent systems a shared way to represent human participation without pretending to read minds or letting a capability estimate determine authority.

The human is already part of the agent network

Agent workflows depend on people when:

  • the model is uncertain or lacks context;
  • policy requires approval;
  • a rare case needs domain expertise;
  • an action has legal, financial, or safety consequences;
  • several agents disagree;
  • automation has degraded;
  • a customer or operator must supply missing information;
  • a person needs to take control.

Most systems record the response but not the runtime episode that produced it. They cannot reliably answer:

  • Did the right person receive the request?
  • Did the necessary context arrive?
  • Was there enough time to inspect the evidence?
  • Did the person actually verify the critical material?
  • Was the handover understood?
  • Did the intervention improve the result?
  • Did the same failure recur?

HRT turns those questions into traceable, evaluable system behavior.

What HRT adds

Human-aware traces

hrt.node, hrt.state, hrt.span, hrt.event, and hrt.trace connect human participation to model, tool, task, and outcome timelines.

Processor discovery

An agent can discover compatible cognitive or human-performance processors, inspect declared capabilities, and request only the authorized metric required for the task.

Better escalation

An agent can select when, how, and to whom a task is escalated using declared policy and current operational evidence. HRT does not grant authority; it helps execute a pre-authorized routing policy with better context.

Handover integrity

The system can record whether critical context was delivered, sampled, acknowledged, and acted upon within its useful window.

Evaluation

Human interventions can be connected to later outcomes. Agent teams can learn whether escalation, timing, interface, processor selection, or evidence presentation actually helped.

Multi-provider resilience

The agent system can select or combine processors by task, modality, latency, jurisdiction, cost, or evidence. Disagreement can trigger review instead of being silently averaged.

A2A, OpenTelemetry, and HRT

HRT does not need to replace agent or telemetry protocols.

  • A2A or another agent transport can advertise HRT-aware skills, negotiate authorization, assign evaluation or inference tasks, and return results.
  • OpenTelemetry or another telemetry path can carry high-volume observations and correlate them with model, tool, and service traces.
  • HRT defines the human-runtime semantics, processor profiles, vertical context, provenance, authorization, and outcome records inside those exchanges.

The agent protocol moves the task. The telemetry system moves the observations. HRT makes the human part interpretable and governable.

What an agent platform gains

  • a reusable human-escalation interface;
  • visibility into handover and approval quality;
  • compatibility with several processors and devices;
  • stronger agent evaluations that include human contribution;
  • auditable provenance for human-derived context;
  • vertical profiles that reduce domain-specific prompting and guesswork;
  • new routing, coaching, assurance, and workflow products;
  • a safer answer to the claim that a person is “in the loop.”

Boundaries that matter

HRT does not let an agent:

  • infer private mental content;
  • convert a provider estimate into a diagnosis;
  • create or remove a person's authority;
  • reuse a temporary state as a permanent profile;
  • access raw biometric data without authorization;
  • treat visible sampling as proof of comprehension;
  • make consequential decisions from an unsupported score;
  • hide which processor or evidence informed an action.

These limits are part of the product value. They make human context safer to integrate across organizations.

A first agent-system project

  1. Choose one workflow where an agent asks a person for help.
  2. Create an HRT span around the escalation and response.
  3. Record the context package, evidence inspected, timing, action, and outcome.
  4. Add one transparent metric or one protected processor output.
  5. Use an A2A extension or ordinary API binding to advertise and invoke the capability.
  6. Compare the workflow with and without the HRT context.
  7. Evaluate whether the human intervention became more timely, informed, or effective.

Concise case for participation

Your agents already depend on people, but the trace usually stops when a person enters the workflow. HRT gives the agent system a structured way to discover authorized human-performance capabilities, coordinate handovers, attach provenance and validity to human-derived context, and evaluate whether an intervention improved the outcome. A2A can carry the task. OpenTelemetry can carry the operational trace. HRT supplies the human-runtime semantics and governance between them. The result is not surveillance or automatic authority. It is a more observable, composable, and responsible human–agent system.