Complete Community Edition under Apache 2.0. No telemetry.See the source
Agents City

How to compare research with AI agents without groupthink

Collect independent positions from literature, methods, data, and ethics before one research chair sees the full comparison.

Reviewed against Agents City 0.3.0-beta.22 · commit c985198 ↗
01

The real problem

A researcher has a literature repository, analysis notebooks, a dataset contract, and ethics notes. Asking agents in one shared chat encourages them to converge before contradictory evidence is visible.

02

Target outcome

A research city holds one question and keeps initial positions isolated until the response barrier completes, then attributes synthesis and verification.

evidence-review

Domain
research
Goal
Assess whether the available evidence supports the proposed conclusion and identify what would falsify it.
Chair seat
research-director
Units and evidence
literatureresearcher

Claims, sources, prior findings, and contradictory studies

analysismethods

Study design, code, assumptions, and reproducibility

data-contractresearch-ops

Provenance, missingness, transformations, and access constraints

ethics-notesethics

Consent, harm, privacy, and publication limits

04

Before you start

  1. 01

    Evidence stored in local repositories or folders that each agent can inspect within its scope.

  2. 02

    A falsifiable research question rather than a request to confirm a preferred conclusion.

  3. 03

    Human review appropriate to the field, especially for health, safety, or policy claims.

05

How to do it

  1. Step 01

    Create a research city

    Choose the research domain and research-director chair. Write the goal around the evidence question and the standard required to support a conclusion.

    TERMINAL
    agents-city cities create evidence-reviewagents-city seat --city evidence-review --domain research --role research-director
  2. Step 02

    Assign evidence repositories and roles

    Select literature, analysis, data, and ethics sources. Give each member a role that matches its evidence and record what lies outside that scope.

    TERMINAL
    agents-city seat --city evidence-review --reposagents-city seat --city evidence-review --agent-roles
  3. Step 03

    Open the review with a response barrier

    Ask for claims, source locations, uncertainty, contradictions, and falsification conditions. Isolated first positions prevent participants from copying the earliest narrative.

    TERMINAL
    agents-city committee open --input evidence-question.json
  4. Step 04

    Synthesise disagreements and verify

    The chair compares evidence without voting. Assign verification to a member that did not produce the chosen analysis and close only after a pass.

    TERMINAL
    agents-city committee show <deliberation-id>
06

What you should get

  • Claims remain attributable to a repository, source, or analysis path.
  • Contradictory evidence and uncertainty survive the synthesis step.
  • The final act records what was decided, what remains unknown, and what would reopen the review.
07

What this does not solve

  • Model output is not evidence unless it points to inspectable source material.
  • Agents City does not validate scientific truth or replace peer review and domain expertise.
  • Sensitive research data still needs appropriate OS, storage, and network isolation.