🗂️

CodeR

Self-ReportedCurated

5-agent supervised pipeline — 28.33% SWE-bench Lite, $3.09/issue.

Huawei Cloud + CAS + SMU + PKU (CodeR)· Operating since Jun 1, 2024· active
Curated from arXiv 2406.01304 — CodeR — 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 · CodeR paper published — arXiv 2406.013042y ago

Spec sheet

The benchmark fields — designed for comparison across teams.

Topology
Supervisor
Agent count
5
Platform
CodeR
Runs on
CodeR ×5
Industries
software-delivery
Task kinds
github-issue-resolution
Trust tier
Self-Reported
Proof entries
1

Topology & roster

Supervisor

Supervisor (Manager) with pre-defined task-graph plans. The Manager selects a plan (encoded as a JSON task graph) and dispatches Reproducer, Fault Localizer, and Editor per the plan; Verifier confirms the final patch; all stages report back to the Manager.

System wiring

Wiring from published architecture

Wiring from published architecture: arXiv 2406.01304 — CodeR (full text)

Node details

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

OrchestratorManager Agent
Tool
Manager Agent
Autonomy
Runs autonomously

Selects a plan (A–D; only A & B used), submits the final patch

View agent profile →
Sends
  • dispatches (plan step) → Reproducer Agent
  • dispatches (plan step) → Fault Localizer Agent
  • dispatches (plan step) → Editor Agent
Receives
  • reports reproduction ← Reproducer Agent
  • reports fault location ← Fault Localizer Agent
  • reports verification result ← Verifier Agent
BuilderEditor Agent
Tool
Editor Agent
Autonomy
Runs autonomously

Iteratively edits code, reusing AutoCodeRover search actions

View agent profile →
Sends
  • submits patch for test → Verifier Agent
Receives
  • dispatches (plan step) ← Manager Agent
BuilderReproducer Agent
Tool
Reproducer Agent
Autonomy
Runs autonomously

Generates reproduce.py to confirm the reported bug

View agent profile →
Sends
  • reports reproduction → Manager Agent
Receives
  • dispatches (plan step) ← Manager Agent
QAVerifier Agent
Tool
Verifier Agent
Autonomy
Runs autonomously

Runs reproduced + integration tests to confirm the fix

View agent profile →
Sends
  • reports verification result → Manager Agent
Receives
  • submits patch for test ← Editor Agent
ResearchFault Localizer Agent
Tool
Fault Localizer Agent
Autonomy
Runs autonomously

Combines SBFL + BM25 scoring to localize the fault

View agent profile →
Sends
  • reports fault location → Manager Agent
Receives
  • dispatches (plan step) ← Manager 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. The orchestrator routes the work

    Manager Agent dispatches (plan step) build work to Reproducer Agent, dispatches (plan step) research/ops work to Fault Localizer Agent, and dispatches (plan step) build work to Editor Agent.

  2. The builders execute

    Reproducer Agent (Generates reproduce.py to confirm the reported bug) and Editor Agent (Iteratively edits code, reusing AutoCodeRover search actions) build the work.

  3. Independent review gates the work

    Verifier Agent reviews the work — Runs reproduced + integration tests to confirm the fix. 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.

Setup order

  1. 1.Stand up the orchestrator: Manager Agent.
  2. 2.Wire Editor Agent: it receives "dispatches (plan step)" from Manager Agent and sends "submits patch for test" to Verifier Agent. Wire Reproducer Agent: it receives "dispatches (plan step)" from Manager Agent and sends "reports reproduction" to Manager Agent. Wire Verifier Agent: it receives "submits patch for test" from Editor Agent and sends "reports verification result" to Manager Agent. Wire Fault Localizer Agent: it receives "dispatches (plan step)" from Manager Agent and sends "reports fault location" to Manager Agent.

Performance metrics

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

SWE-bench Lite pass@1 (single submission)
28.3%
evidence-linked

85/300 resolved (one submission per issue). Source: arXiv 2406.01304. The paper's own Results prose separately states 'addressing 84 of 300' — an internal inconsistency in the paper itself (85 elsewhere, incl. the abstract and Table 4). [evidence_linked]

as of Jun 1, 2024

Token economics

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

Avg cost per issue
$3.09

≈299K tokens / $3.09 per issue avg; 30.39 API requests avg; hard cap $8/issue. Source: arXiv 2406.01304. [evidence_linked]

as of Jun 1, 2024

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

The ablation is the strongest evidence here: removing the multi-agent structure and task graph collapses performance from 22% to 10% on the same 50-issue subset — the paper's own data shows the coordination structure, not just the underlying model, drives most of the result.

How it was built

Custom orchestration reusing search actions from SWE-agent and AutoCodeRover; plans encoded as JSON task graphs and parsed into an executable graph. Structured inter-agent reports only, no free-form chatter. Open-source at github.com/NL2Code/CodeR, but with no LICENSE file.

Oversight model

None in benchmark evaluation — the Manager autonomously selects a plan and coordinates the other four agents to a single submitted patch per issue.

Proof (1)

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

  1. ArtifactJun 1, 2024evidence-linked

    CodeR paper published — arXiv 2406.01304

    5-agent supervised pipeline: 28.33% (85/300) SWE-bench Lite, ~$3.09/issue avg. Ablation: full CodeR 22%, without multi-agent + task graph 10%, without fault localization 14% (50-issue subset). No LICENSE file exists in the repository.

    https://arxiv.org/abs/2406.01304

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