# AgenticConsensus

> AgenticConsensus is a shared, versioned project state for human–AI coding collaboration.

Consensus Control is the project's open-source local review workspace.

## Public resources

- [Project website](https://tianfu.wang/AgenticConsensus/): overview, illustrative walkthrough, real workspace captures, paper, and setup.
- [arXiv preprint](https://arxiv.org/abs/2604.17883)
- [alphaXiv reader](https://www.alphaxiv.org/abs/2604.17883)
- [Implementation](https://github.com/GeminiLight/AgenticConsensus): Python and JavaScript, Apache-2.0.
- [Website source](https://github.com/GeminiLight/AgenticConsensus/tree/main/website)
- [Getting started](https://github.com/GeminiLight/AgenticConsensus/blob/main/docs/usage.md)
- [Architecture](https://github.com/GeminiLight/AgenticConsensus/blob/main/docs/architecture.md)
- [Citation metadata](https://github.com/GeminiLight/AgenticConsensus/blob/main/CITATION.cff)
- [Tianfu Wang's publication page](https://tianfu.wang/publications/#consensus-layer): personal-site entry linking the paper, project website, and code.

## Paper

Title: Scaling Human-AI Coding Collaboration Requires a Governable Consensus Layer
Authors: Tianfu Wang, Zhezheng Hao, Yin Wu, Wei Wu, Qiang Lin, Hande Dong, Nicholas Jing Yuan, Hui Xiong.
Current project affiliations: The Hong Kong University of Science and Technology (Guangzhou), Zhejiang University, Tencent. The earlier arXiv version lists its own publication-time affiliations.
Year: 2026. arXiv:2604.17883 [cs.SE], first submitted 20 April 2026.
DOI: https://doi.org/10.48550/arXiv.2604.17883
Version note: This website describes the current research and implementation. The linked arXiv preprint is an earlier version; the authors are preparing an update. No conference acceptance is asserted.

## Core idea

AgenticConsensus connects selected intent, structural commitments, executable artifacts, evidence, and authorized decisions in a versioned project state. Consensus means the currently authorized state, not agent voting or a guarantee of truth. Contested claims, uncertainties, and scoped waivers remain explicit.

- Intent (I): the original request, alternatives, clarification, and selection.
- Commitments (C): interfaces, dependencies, invariants, and authority policies.
- Artifacts (A): exact code, configuration, and revisions.
- Evidence (E): checks linked to claims and artifact versions, with scope and freshness.
- Realization (Phi) links a selected structural change to candidate artifacts.
- Rehydration (Psi) reads artifact changes back into the model and records discrepancies and unobservable state.

Project-state review applies to shared interfaces, authority or policy changes, compatibility windows, and evidence obligations beyond the diff. Local edits can stay on ordinary diff-and-test review.

## What is implemented

Consensus Control is an authenticated local review workspace with read-only GitHub pull-request import, commitment and source inspection, a change map, evidence obligations, authorized decisions, and an audit history. AI assistance is optional and advisory. Automated source analysis covers Python imports. The executable realization path is bounded to a single Python import substitution; it is not general code synthesis.

The GitHub connector does not push, merge, create, or delete pull requests. A local governed decision is not a GitHub merge. Deployment is localhost or behind operator-managed TLS; this is not a hosted multi-tenant SaaS.

## Research status

The planned comparative evaluation separates native practice (B0), a governance ledger (B1), and equivalent relational review (B2). Controlled causal benefits and the incremental value of graph navigation remain future evaluation questions.

The homepage capability-migration walkthrough is illustrative. Workspace screenshots are recorded system captures. A source reading is not behavioral verification, and recorded evidence does not automatically satisfy an obligation.

## Citation

```bibtex
@misc{wang2026scalinghumanaicodingcollaboration,
  title = {Scaling Human-AI Coding Collaboration Requires a Governable Consensus Layer},
  author = {Wang, Tianfu and Hao, Zhezheng and Wu, Yin and Wu, Wei and Lin, Qiang and Dong, Hande and Yuan, Nicholas Jing and Xiong, Hui},
  year = {2026},
  eprint = {2604.17883},
  archivePrefix = {arXiv},
  doi = {10.48550/arXiv.2604.17883},
  url = {https://arxiv.org/abs/2604.17883}
}
```