Collaborative Human-Agent Protocol (CHAP) | AI Research

Collaborative Human-Agent Protocol (CHAP) addresses the growing challenge of managing work where humans and AI agents collaborate in operational roles. As foundation models move beyond simple chat responses to tasks like drafting contracts, managing claims, or writing code, current systems often fail to track the "moment of human judgment." CHAP provides a standardized, auditable framework for these interactions, ensuring that every decision, edit, and approval is recorded in a structured, replayable format rather than being lost in fragmented chat threads or internal code.

The Problem: The Collaboration Gap

In modern production environments, work is increasingly distributed across teams, time zones, and different AI systems. While existing standards like the Model Context Protocol (MCP) help agents access tools and data, and Agent-to-Agent (A2A) protocols help agents communicate, there is no standard for the "shared workspace" where humans and agents actually perform accountable work together. Currently, when a human overrides an agent’s draft or approves a high-stakes decision, these actions are often buried in application-specific logs or tribal memory, making it impossible to audit the process later or understand why a specific decision was made.

How CHAP Works

CHAP functions as a collaboration layer that sits alongside existing infrastructure. It is built on a "Core plus profiles" architecture:

• The Core: This provides the essential building blocks for any workspace, including definitions for participants (humans or agents), tasks, artifacts, and an append-only evidence log. It uses a structured envelope model to exchange events, ensuring that every action—such as a task assignment, a human override, or a shift handoff—is captured as a formal, traceable event.

• Composable Profiles: To handle more complex needs, users can add optional profiles. These include features like identity verification, digital signatures, review workflows, and transparency-backed audit logs. This modular design allows teams to start with a simple setup and add governance features only as their specific regulatory or operational requirements demand.

Key Benefits for Accountable Work

By standardizing these collaboration events, CHAP transforms how organizations manage AI-assisted work. A human override is no longer just a text change; it becomes a structured event containing a diff, a rationale, and a content hash. Similarly, handoffs between shifts become portable envelopes, and human approvals become non-repudiable, signed decisions. This creates a permanent, verifiable evidence chain that can be replayed years later, which is critical for industries like healthcare, finance, and legal services where accountability and regulatory compliance are mandatory.

Scope and Limitations

CHAP is designed to be narrow and focused. It does not attempt to replace existing workflow engines, identity providers, or policy systems; instead, it is intended to compose with them. The protocol deliberately avoids defining domain-specific taxonomies or regulatory requirements, leaving those to the deploying organization. It provides the infrastructure for a "Human-Agent Symphony," but the sufficiency of the evidence recorded—and whether it meets specific legal or safety standards—remains a socio-technical determination for the users implementing the protocol.