Skip to content

EvoMaster: A Foundational Evolving Agent Framework for Agentic Science at Scale

🕒 Published (v1): 2026-04-19 12:26 UTC · Source: Arxiv · link

Why this paper was selected

Foundational evolving agent framework enabling iterative agentic science at scale

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

EvoMaster is a domain-agnostic foundational agent framework for autonomous scientific research that encodes continuous self-evolution—iterative hypothesis refinement, self-critique, and cross-cycle knowledge accumulation—directly into its execution architecture. New domain-specific scientific agents can be built atop it in ~100 lines of code. On four benchmarks (HLE, MLE-Bench Lite, BrowseComp, FrontierScience) it achieves 41.1%, 75.8%, 73.3%, and 53.3%, outperforming the general-purpose OpenClaw baseline by +159% to +316%.

Problem

Existing scientific agent systems are monolithic and domain-siloed, causing massive duplication of harness-level engineering (tool orchestration, trajectory management, error recovery) across disciplines. They also follow a single-pass execution paradigm: agents run once and terminate, with no mechanism to accumulate knowledge or refine strategies across experimental cycles—contradicting the iterative nature of the scientific method. There is no foundational abstraction layer that enables horizontal scaling of agentic capabilities across scientific domains.

Method

EvoMaster separates concerns across three orthogonal execution layers:

  1. Playground (Orchestration): Declares multi-agent topologies via AgentSlots and the @register_playground decorator, supporting sequential hand-offs, parallel exploration, and peer-review cycles.
  2. Exp (Experiment Execution): Manages single-experiment lifecycle with YAML-based configuration manifests and a thread-safe trajectory system (structured JSON logging of every turn, tool invocation, and token statistic).
  3. Agent Engine: Drives a multi-turn reactive loop—reason → invoke tools → observe → self-critique—sustained over hundreds of turns via a ContextManager with dynamic LLM-based summarization and sliding-window context compression.

The Capability Layer provides a universal interface: a Tool System using the Action-Execution-Observation pattern with native MCP integration; a Skill System that injects domain knowledge hierarchically (metadata in-context, instructions loaded on-demand); and an LLM abstraction standardizing responses across 100+ models via LiteLLM. New domains plug in by defining a new Playground, leaving core execution untouched. The SciMaster ecosystem (ML-Master 2.0, X-Master, Browse-Master, PhysMaster, EmboMaster) is built on this shared base.

Key Contributions

  • EvoMaster framework: A domain-agnostic foundational harness encoding iterative self-evolution (multi-turn reactive loop + context management) as a first-class primitive, deployable for new domains in ~100 lines of code.
  • SciMaster ecosystem: Six domain-specific agents (ML, general science, web retrieval, physics, embodied intelligence) built on EvoMaster, demonstrating rapid horizontal scaling via the shared harness.
  • Experiment-ready observability: YAML configuration manifests + structured JSON trajectory system providing full reproducibility and auditability for long-horizon agent experiments.
  • Benchmark validation: State-of-the-art results on HLE, MLE-Bench Lite, BrowseComp, and FrontierScience against a strong open-source general-purpose baseline (OpenClaw, GPT-5.4 backend for both).

Results

  • HLE: 41.1% (EvoMaster) vs. 13.6% (OpenClaw), +202% relative.
  • MLE-Bench Lite: 75.8% medal rate (EvoMaster) vs. 18.2% (OpenClaw), +316% relative; medals on ~17/22 Kaggle tasks vs. 4/18 for OpenClaw; also outperforms MLE-STAR-Pro-1.5 (68.18%) and R&D-Agent (68.18%) on Any Medal metric.
  • BrowseComp: 73.3% (EvoMaster) vs. 28.3% (OpenClaw), +159% relative; Map + Search subcategory 100.0% vs. 25.0%.
  • FrontierScience: 53.3% (EvoMaster) vs. 18.3% (OpenClaw), +191% relative.
  • MLE-Bench performance improves progressively over time within a single run, empirically demonstrating the self-evolution loop.

Limitations

  • Comparisons are made against a single general-purpose baseline (OpenClaw); no ablation isolating the contribution of the evolution loop vs. domain-specific SciMaster agent design.
  • All experiments use GPT-5.4 as the backend; generalizability to weaker or open-weight models is not assessed.
  • The ~100 lines of code claim for new domain onboarding is stated but not formally measured or reproducibly demonstrated in the text.
  • Several SciMaster agents (X-Master 2.0, PhysMaster, EmboMaster) are listed as "coming soon" and not publicly available, limiting ecosystem reproducibility.
  • No analysis of failure modes or cases where the self-evolution loop degrades (e.g., compounding errors over hundreds of turns).

Relevance to Harnesses / Meta-Harnesses

EvoMaster is explicitly architected as a meta-harness—a domain-agnostic foundational substrate whose primary value proposition is eliminating duplicated harness engineering across domain-specific agents. Its three-layer decoupling (Playground / Exp / Agent) and the "experiment harness as first-class primitive" principle (YAML manifests, JSON trajectory logging) represent a direct design response to the lack of a shared abstraction layer in scientific agent development. The framework's plugin model (MCP tool integration, hierarchical Skill injection, unified LLM abstraction) is a concrete template for how meta-harnesses can expose extension points without coupling domain logic to execution logic. For researchers tracking harness design patterns, EvoMaster demonstrates that embedding the iterative/evolutionary loop at the harness level—rather than in domain-specific code—yields large, cross-domain performance gains.