Intent
The request and its agreed interpretation.
Human + agent
collaborationResearch & open-source software / 2026
Build with agents.
Stay in control.
See what an agent’s change means for your project. Connect the intent, code, and evidence. Make an informed decision.
AgenticConsensus is a shared, versioned project state for human–AI coding collaboration.
One change. The whole context.
The intentTask-scoped accessAgreed
The agent’s changeWorkspace-wide tokenMismatch
The evidenceTests pass. A check is missing.
The human decisionResolve the gap before approval.
Illustrative scenario Explore any step ↗
Behind the project
Built on research.
Made for collaboration.
01 / A closer look
The tests pass.
Does the change belong?
A working connector quietly broadens access. Follow the mismatch from the original request to the review decision.
Consensus Control /
Capability migrationIllustrative walkthrough
“Use task-scoped access.
Keep the existing connectors working.”
C01
The shared commitment
Use only the capabilities granted to the current task.
Legacy exceptions need a scope and an expiry.
Keep the interpretation and who agreed on record.
A small import.
A bigger change in authority.
connector.py Illustrative diff
− from task_auth import get_token+ from workspace_auth import get_token
ConnectorWorkspace tokenTask scope
The new dependency may violate task scope. Dynamic credential resolution can remain unobservable.
Passing tests.
Unanswered questions.
Unit testsMocks allow every toolPass · limited scope
Runtime access traceRecorded against an older revisionStale
Denied-capability checkRequired for the proposed revisionMissing
Evidence must support this claim, on this version—not just show a passing result.
Agents propose.
Humans authorize.
Not ready to approve
Resolve the access-policy mismatch and supply current evidence.
* Only where policy permits. Every decision records its actor, rationale, scope, and exact project version.
Illustrative capability-migration scenario. No live repository actions or measured results.
02 / The consensus layer
Keep the context.
Across every change.
Prompts get buried. Code keeps changing. A shared project state preserves the decisions that connect them.
Commitments
The interfaces, dependencies, and rules to preserve.
Artifacts
The code and configuration at a specific revision.
Evidence
What was checked, for which claim and version.
Consensus means an authorized project state. Disagreements, uncertainty, and scoped exceptions remain visible.
Context travels both ways. Agreed changes guide implementation; changed artifacts reveal conflicts and stale evidence.
03 / The working system
Meet Consensus Control.
The open-source review workspace. Inspect a commitment, follow it into the code, and see what still needs evidence.
An agreed commitment, beside the code. An English-only synthetic demo. Linked files are marked “Not interpreted” until a source reading is recorded.
The decision follows the evidence
Three claims still need evidence. Approval stays blocked until the requirements are resolved.
Know what blocks approval. An AI answer can inform a decision; it cannot authorize one.
Local-first. Read-only GitHub import. Optional AI. Captures from the development workspace. See the FAQ for implementation scope.
Run the workspace04 / The research
Scaling Human-AI Coding Collaboration Requires a Governable Consensus Layer.
Our position: more agent-written code needs a shared account of what the project means—and who can change it.
1 HKUST (Guangzhou) 2 Zhejiang
University 3 Tencent
* Corresponding authors
arXiv:2604.17883 · First submitted 20 April 2026. Updated version coming soon. This site follows our newer research and implementation.
Current research summary
Code records what was built. It does not, by itself, explain the choices a project has agreed to preserve. AgenticConsensus links intent, commitments, artifacts, and evidence in a versioned project state that people and agents can inspect and revise.
Agreed changes guide implementation. Updated artifacts reveal conflicts, gaps, and stale evidence. Consensus Control puts this approach into a review workspace, with humans retaining decision authority.
05 / Try it on your machine
Your project.
Your review desk.
Try the demo, import a GitHub PR, or compare local Git revisions.
Python 3.11+Gituv
Full setup & configuration
git clone https://github.com/GeminiLight/AgenticConsensus.git
cd AgenticConsensus
uv venv .venv
uv pip install --python .venv/bin/python -e '.[dev]'
.venv/bin/agentic-consensus demo
.venv/bin/agentic-consensus serve
Open
For the seeded demo, use audit identity
http://127.0.0.1:8765 and create the local
administrator.For the seeded demo, use audit identity
[email protected].
06 / Go deeper
One project. Open paths in.
Read it. Run it. Build on it.
01The paperThe public preprint · updated version coming soonarXiv:2604.1788302The implementationConsensus Control · source, setup, and contributionsGitHub / Apache-2.003The project websiteThis page · design, walkthrough, and public resourcesGitHub04The author's publication pageTianfu Wang · related research, paper, website, and codePersonal website
Cite the arXiv preprint
View 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}
}A few useful distinctions
Before you dive in.
Is consensus a vote between agents?
No. Consensus is the authorized project state, including unresolved claims and scoped exceptions. Agent agreement alone does not establish truth or grant permission.
How does this relate to code review?
It connects a diff to the commitments it affects, the intent behind it, and the evidence that still applies. Your existing code review gains a shared project context.
What can I run today?
Consensus Control supports local review, read-only GitHub PR import, source inspection, evidence requirements, and authorized decisions. Automated analysis covers Python imports; executable changes are currently limited to a single import substitution.
Does approval merge my GitHub pull request?
No. Approval updates local project state and its audit history. The read-only GitHub connection cannot push code or merge a PR.
Move the code forward.
Keep the meaning with it.