Eureka

Eureka is a task-conditioned Meta-Agent architecture for long-horizon scientific discovery. This repository contains the paper-aligned public implementation of the orchestration runtime, task-conditioned specialized-agent components, schemas, and reproducibility utilities.

Product architecture note. Eureka is a research sub-architecture of Tanglang, developed by ManXis.

Paper: Eureka: Task-Conditioned Meta-Agent Orchestration for Scientific Discovery
Repository: https://github.com/manxis-contact/Eureka

Authors

1. Alizer Wong (corresponding author) — ManXis
2. Heng Cui — ManXis
3. Yi Tan — School of Information Engineering, Guangdong University of Technology
4. Xiongchao Zhan — School of Automation, Guangdong University of Technology
5. Liang Lin — School of Artificial Intelligence, South China Normal University
6. Yuxiang Guo — Shanghai Jiao Tong University
7. Zhaorong Dai — Pratt School of Engineering, Duke University
8. Zixin Zeng — School of Computer Science and Technology, Guangdong University of Technology
9. Wenyuan Li — Hokkaido University

Correspondence: contact@manxis.org

Repository Layout

src/eureka/meta_agent/      Dynamic obligation orchestration, promotion, evolution, and control schemas
src/eureka/theory_agent/    Theory-discovery policies, verification, memory, and incremental-runtime specifications
src/eureka/math_agent/      Open-conjecture runtime, typed fact/claim IR, exact primitives, retrieval, and verifier routing
tests/                      Deterministic unit tests for core public components
paper/                      Paper source, bibliography, and report PDF
assets/                     ManXis logo and Eureka framework figure

Core Components

• Dynamic obligation orchestration: receding-horizon planning, ready-frontier backpressure, typed plan deltas, and event-driven control.
• Task-conditioned Macro-Agent formation: architecture promotion based on state sharing, dependency density, recurrence, continuity, and amortized execution cost.
• Governed self-evolution: bounded architecture adaptation with diagnosis, cost-aware admission, immutable contracts, and rollback.
• Theory-Discovery Agent: hypothesis search, counterexample pressure, assumption auditing, structured memory, verification ladders, and incremental evidence processing.
• Math/Conjecture Agent: primary-source intake, typed fact/claim/proof IR, exact structural primitives, incremental mining, and verifier routing.
• Compiled scientific runtime: dependency-local rebuilds, certificate-aware reuse, context paging, and isolated lease merging.

Quick Start

The public package uses only the Python standard library for the core runtime.

python -m pytest -q

A minimal orchestration example:

from eureka.meta_agent.runtime.hotspot import HotspotFeatures, should_promote

features = HotspotFeatures(
    nodes=12,
    state_sharing=0.8,
    dependency_density=0.7,
    operator_recurrence=0.7,
    verifier_recurrence=0.6,
    continuity_need=0.8,
    horizon=0.9,
    repeated_plan_cost=40,
    repeated_context_cost=30,
    coordination_cost=20,
    expected_rebase_cost=10,
    lazy_compile_cost=35,
    local_runtime_cost=25,
)

print(should_promote(features))

The open-conjecture runtime can be invoked from Python:

from eureka.math_agent.tools.runtime import run
print(run("Study a specified open conjecture", []))

Paper Assets

The repository includes the ManXis team logo and the paper framework figure under assets/. The paper source references the architecture as Eureka; Tanglang is identified only as the product-level system that contains Eureka as a sub-architecture.