← Selected work

Case study · Multi-Agent product

牛来生米小队

An in-development market-event application where four product Agents share evidence and chronology but retain different responsibilities, votes and action permissions.

Status
Level 2 evidence · freeze pending
My role
Product rules, permissions & UAT
System
Four Agents · one timeline
Audit snapshot
352 / 352 automated tests
云雀牛来One room · One Case · One timeline牛来妈妈豹拉

The product problem

A multi-Agent interface can easily become four disconnected chat windows. This project instead treats roles, votes, permissions and evidence as product rules. Candidate evidence, discussion, decisions, paper results, content drafts and outcomes remain addressable from one event history.

I designed the role split around the actual project need rather than using one fixed Agent template. I remain responsible for product direction, permission boundaries, key trade-offs, failure review and full-flow acceptance.

Four roles

云雀

Read only

Finds public hotspot candidates and records provenance. It can nominate; it cannot decide or execute alone.

牛来

Paper only

Consumes an eligible structured decision and writes the paper result back to the timeline only after the configured threshold.

牛来妈妈

Draft gated

Produces X content from canonical events. Publishing is isolated from ordinary model work and requires explicit lifecycle authority.

豹拉

Draft only

Creates Binance Square drafts from the same evidence trail. Drafting does not become login, posting or execution authority.

System spine

Each stage records what happened and what authority was available. Facts, inferences and open questions remain distinguishable; decisions require the configured threshold; actions remain permission-gated; unsafe failure paths close rather than silently continue.

EvidenceFact · inference · open
→
DecisionIndependent roles · 3-of-4 threshold
→
ActionPermission-gated record

What is already evidenced

  1. 01

    Permissions are explicit

    Model work, data reading, drafting, publishing and controlled execution are separated rather than inherited from a generic Agent role.

  2. 02

    Failures become product records

    Model fallback, state leakage, identity collision, stale deployment output and terminal-sync gaps are documented instead of hidden.

  3. 03

    Verification has a dated boundary

    The 2026-08-25 repository audit recorded 352/352 automated tests passing plus the complete offline validation command.

  4. 04

    Owner UAT is full-flow and iterative

    I ran the complete flow, found detail issues after Agent implementations, directed corrections and re-verified the results.

Contribution

What I owned

  • Product purpose and role design
  • Voting and permission rules
  • Model, API and workflow trade-offs
  • Failure review, corrections and acceptance

What Agents accelerated

  • Substantial implementation and tests
  • Integration, evidence documents and regression checks
  • Deployment tasks under project rules
  • Draft product outputs within configured permissions

Status and boundary

This is an in-development Level 2 evidence case.

A formal release freeze, unified cost ledger, some recovery-path closure and external user or business outcomes remain pending.

The project is an unofficial fan-parody work and does not claim an official relationship, cooperation, sponsorship or endorsement. Public visuals and deeper transaction evidence remain subject to separate IP and privacy review.

Current flagship

See the completed core-playability case.

FightGame remains the first portfolio case until this project reaches its next evidence milestone.

View FightGameAll work