OpenWorkProof: Verifiable Work Contracts for AI Agent Systems

Introduction

AI agents are becoming increasingly capable, yet verifying their actions and ensuring compliance with human intent remains a significant challenge. OpenWorkProof tackles this by providing an open protocol for AI agent work contracts and verifiable execution. It aims to make every agent action auditable and subject to human judgment, ensuring clarity on who authorized a task, what the agent actually did, and whether it stayed within agreed boundaries.

Why Use It & Key Benefits

OpenWorkProof transforms an agent's work into a portable, verifiable chain of evidence. It addresses crucial questions that traditional logs often miss, such as:

• Who authorized the agent's actions?
• Did the agent use only permitted tools and resources?
• Are the outputs, like patches or reports, genuinely from the claimed execution?
• Can a third party independently review the facts without accessing the original systems?

Key benefits and features include:

• Verifiable Evidence Packages: Each work instance can be exported as a signed evidence package, allowing independent verification by third parties using only the package and public keys.
• Six Connected Facts: The protocol links work purpose, signed authorization, pre-execution decisions, action receipts, independent verification, and human acceptance.
• Separation of Roles: Distinct roles like Maintainer, Manager, Developer, Verifier, Sidecar, and Acceptor prevent a single agent from acting as executor, verifier, and final acceptor.
• Human Agency Profile: This feature ensures that human decision-making remains central, allowing humans to define what capabilities agents can autonomously use, what requires human review, and what is denied.
• Verification and Acceptance Separation: OpenWorkProof explicitly separates the technical verification of evidence from the human decision of accepting or rejecting the work, preventing "verified" from being conflated with "accepted" or "paid."
• Verification Integrity: Mechanisms are in place to ensure that verification results themselves are trustworthy, checking for population capture and control failure signatures.

Installation

You can easily install OpenWorkProof using pip:

python -m pip install openworkproof
owp --help

For development or to modify the protocol, you can install from source:

git clone https://github.com/dengyier/OpenWorkProof.git
cd OpenWorkProof
python -m venv .venv
source .venv/bin/activate
python -m pip install -e .
owp --help

Examples

To see the Human Agency Profile in action, run the minimal example:

python examples/human_agency_profile_v01.py

Expected output will include:

profile verified  : True
resolved status   : active
owp.repo_read     : delegated -> allowed
owp.apply_patch   : reserved -> AGENCY_HUMAN_DECISION_REQUIRED

You can also verify existing bundles:

# For Surface Bundle
owp surface-verify PATH

# For Acceptance Bundle
owp acceptance-bundle-verify DIRECTORY

More details on offline verification are available in the official documentation.

Links

• GitHub Repository: https://github.com/dengyier/OpenWorkProof
• PyPI: https://pypi.org/project/openworkproof/
• Protocol Documentation: https://github.com/dengyier/OpenWorkProof/blob/main/docs/protocol/human-agency-profile-v0.1.md
• Offline Verification Guide: https://github.com/dengyier/OpenWorkProof/blob/main/docs/offline-verification.md