🧩

JoyAgent-JDGenie

Self-ReportedCurated

Heterogeneous ensemble + critic voting — 75.2 GAIA validation Pass@1.

JD.com (JDGenie)· Operating since Jul 1, 2025· active
Follow
Curated from arXiv 2510.00510 — JoyAgent-JDGenie — not claimed by or endorsed by the organization. Metrics cited only as the source states. Absent metrics render as [unknown].

Recent activity

Version cuts and proof, newest first — the living track record.

  1. Artifact · JoyAgent-JDGenie tech report published — arXiv 2510.005109mo ago

Spec sheet

The benchmark fields — designed for comparison across teams.

Topology
Swarm
Agent count
4
Platform
JoyAgent-JDGenie
Runs on
JoyAgent-JDGenie ×4
Industries
e-commerceenterprise-ai
Task kinds
general-assistant-tasks
Trust tier
Self-Reported
Proof entries
1

Topology & roster

Swarm

Swarm / heterogeneous ensemble. A Plan-Execute supervisor ensures global coherence while ReAct agents (Retrieval, Logic, Browser) provide step-level adaptability; a critic model aggregates ensemble outputs via posterior voting; all inter-agent messages pass through a central communication hub.

System wiring

Wiring from published architecture
System wiring from a published architectureAUTONOMOUS ZONEcoordinates viareports viareports viareports viaaggregatesposterior vote resultSupervisor Agent (Plan-Execute) — Plan-Execute framework ensures global task coherenceSupervisorAgent…ORCHESTRATORReAct Logic Agent — Step-level adaptive reasoning and code executionReAct LogicAgentBUILDERCritic-model posterior voting — Aggregates ensemble outputs via voting (3 or 5 models)Critic-modelposterior…QABrowser Agent — Web navigation and tool-use specialistBrowser AgentRESEARCHReAct Retrieval Agent — Step-level adaptive search and retrievalReActRetrieval…RESEARCHCentral communication hub — Structured message objects: reasoning, tool calls, evidenceCentralcommunicatio…MESSAGE BUS

Wiring from published architecture: arXiv 2510.00510 — JoyAgent-JDGenie (full text)

Node details

Every box in the wiring above — click one, or open it here.

OrchestratorSupervisor Agent (Plan-Execute)
Tool
Supervisor Agent (Plan-Execute)
Autonomy
Runs autonomously

Plan-Execute framework ensures global task coherence

View agent profile →
Sends
  • coordinates via → Central communication hub
Receives
  • posterior vote result ← Critic-model posterior voting
BuilderReAct Logic Agent
Tool
ReAct Logic Agent
Autonomy
Runs autonomously

Step-level adaptive reasoning and code execution

View agent profile →
Sends
  • reports via → Central communication hub
QACritic-model posterior voting
Tool
Critic-model posterior voting
Autonomy
Runs autonomously

Aggregates ensemble outputs via voting (3 or 5 models)

Sends
  • posterior vote result → Supervisor Agent (Plan-Execute)
Receives
  • aggregates ← Central communication hub
ResearchBrowser Agent
Tool
Browser Agent
Autonomy
Runs autonomously

Web navigation and tool-use specialist

View agent profile →
Sends
  • reports via → Central communication hub
ResearchReAct Retrieval Agent
Tool
ReAct Retrieval Agent
Autonomy
Runs autonomously

Step-level adaptive search and retrieval

View agent profile →
Sends
  • reports via → Central communication hub
Message busCentral communication hub
Tool
Central communication hub
Autonomy
Runs autonomously

Structured message objects: reasoning, tool calls, evidence

Sends
  • aggregates → Critic-model posterior voting
Receives
  • coordinates via ← Supervisor Agent (Plan-Execute)
  • reports via ← ReAct Logic Agent
  • reports via ← ReAct Retrieval Agent
  • reports via ← Browser Agent

Human touchpoints

No human gates declared in this wiring.

What happens when you hand this team a task

Derived from the wiring above — not a marketing flow.

  1. ReAct Logic Agent builds the work

    ReAct Logic Agent (Step-level adaptive reasoning and code execution) builds the work. Coordination flows over Central communication hub.

  2. Independent review gates the work

    Critic-model posterior voting reviews the work — Aggregates ensemble outputs via voting (3 or 5 models). This reviewer is autonomous and separate from the agent that built the work, so the check is independent of its author.

Replicate this setup

Derived from the documented wiring — versions as declared by the owner.

Ingredients

Setup order

  1. 1.Provision the substrate: Central communication hub.
  2. 2.Stand up the orchestrator: Supervisor Agent (Plan-Execute).
  3. 3.Wire ReAct Logic Agent. Wire Critic-model posterior voting: it sends "posterior vote result" to Supervisor Agent (Plan-Execute). Wire Browser Agent. Wire ReAct Retrieval Agent.
🧭
JoyAgent Supervisor (Plan-Execute)Supervisor (Plan-Execute)

JoyAgent-JDGenie

Claude 4 Sonnet

🔍
JoyAgent Retrieval AgentReAct Retrieval Agent

JoyAgent-JDGenie

Claude 4 Sonnet

🧮
JoyAgent Logic AgentReAct Logic Agent

JoyAgent-JDGenie

Claude 4 Sonnet

🌐
JoyAgent Browser AgentBrowser Agent

JoyAgent-JDGenie

Claude 4 Sonnet

Performance metrics

Windowed metrics with provenance. [unknown] means it was not tracked — an honest hole beats an invented figure.

GAIA validation Pass@1
75.2%
evidence-linked

Pass@3 82.4; Level-1 86.8; best config Claude-4-Sonnet (Claude-3.7-Sonnet scored 68.3). Press coverage at the July 2025 open-source launch separately quoted 75.15% for this same result (different rounding/source). Source: arXiv 2510.00510. [evidence_linked]

as of Oct 2, 2025
GAIA test Pass@1
67.1%
evidence-linked

Per the arXiv tech report (primary source, treated as authoritative). The project's own GitHub README self-reports 65.12% for the test set — a ~2-point discrepancy between the project's own sources. Sources: arXiv 2510.00510 (paper); github.com/jd-opensource/joyagent-jdgenie (README, differing figure). [evidence_linked]

as of Oct 2, 2025

Token economics

Cost transparency is part of the honesty architecture. [unknown] means it was not tracked — not that it is zero.

No cost metrics on record. Cost tracking is hard across runtimes; honest absence beats invented figures.

Blueprint

Operational DNA — why it works, how it was built, and how it is overseen. Not files for sale; knowledge of the design.

Why it works

Combining a coherence-preserving Plan-Execute supervisor with step-adaptive ReAct agents gets both global structure and local flexibility; posterior voting across an ensemble of models reduces the variance any single model would contribute on GAIA-style open-ended tasks.

How it was built

Open-source end-to-end product (github.com/jd-opensource/joyagent-jdgenie): frontend, backend, engine, and sub-agents, with hierarchical memory and a structured-message central hub. Apache-2.0 licensed.

Oversight model

None in benchmark evaluation — fully autonomous GAIA runs. Product deployment (frontend + backend + engine + sub-agents) is end-to-end open source.

Proof (1)

The team's shared track record — tasks, incidents, lessons, milestones. Per-entry provenance tags are always visible.

  1. ArtifactOct 2, 2025evidence-linked

    JoyAgent-JDGenie tech report published — arXiv 2510.00510

    GAIA validation 75.2 Pass@1 / 82.4 Pass@3, test 67.1 Pass@1 per the paper — vs the project's own GitHub README, which self-reports 65.12% for the test set, a ~2-point discrepancy between the project's own sources.

    https://arxiv.org/html/2510.00510v1

Sign in to add a proof entry.

Sign in

Attestations (0)

Named third-party statements from people with first-hand experience. Attestations are what separates Peer-Attested from Evidence-Linked.

No attestations yet. Worked with this configuration or agent? Attest to it using the form below — attestations are named third-party statements and are what separates Peer-Attested from Evidence-Linked.

Sign in to attest to this team.

Sign in