Autonomous AI agents for end-to-end scientific discovery represent a paradigm shift in how research is conducted. These systems leverage large language models (LLMs) coordinated with specialized tools to perform hypothesis generation, experiment design, code execution, result analysis, and even automated paper writing — all with minimal human intervention.

The concept of AI scientist agents emerged from the convergence of LLM capabilities with laboratory automation and evolutionary search. Unlike traditional AI-assisted research tools that handle isolated subtasks, AI scientist agents operate across the full research lifecycle. They maintain persistent memory of prior experiments, adapt strategies based on accumulated results, and coordinate multiple specialized sub-agents to tackle complex scientific problems.

The field accelerated rapidly in 2025-2026 with systems like Sakana AI Scientist, EvoScientist, and InternAgent demonstrating that autonomous agents can produce novel research contributions across domains including machine learning, chemistry, biology, and materials science.

Sakana AI developed one of the first comprehensive AI scientist frameworks. The AI Scientist (versions 1 and 2) autonomously generates research ideas, writes experiment code, runs simulations, analyzes results, and produces full scientific manuscripts. The system handles literature review, hypothesis generation, and data analysis independently, representing a milestone in end-to-end research automation.

EvoScientist introduces an evolving multi-agent framework that continuously improves its research strategies through persistent memory and self-evolution. The architecture comprises three specialized agents:

• Researcher Agent (RA) — responsible for scientific idea generation
• Engineer Agent (EA) — handles experiment implementation and execution
• Evolution Manager Agent (EMA) — distills insights from prior interactions into reusable knowledge

The system maintains two persistent memory modules: an ideation memory that tracks feasible and unsuccessful research directions, and an experimentation memory that captures effective data processing and model training strategies. Experiments show EvoScientist outperforms seven state-of-the-art systems in novelty, feasibility, relevance, and clarity.

InternAgent 1.5 provides a unified agentic framework for long-horizon autonomous scientific discovery across computational and empirical domains. Its architecture includes subsystems for generation, verification, and evolution, supported by deep research capabilities, solution optimization, and long-horizon memory. It achieves leading performance on GAIA, HLE, GPQA, and FrontierScience benchmarks.

AI scientist agents generate hypotheses through data-driven analysis of literature and databases. Evolutionary approaches like those in EvoScientist and AlphaEvolve couple LLMs with evolutionary search to propose, test, and refine hypotheses iteratively. AlphaEvolve demonstrated this by discovering a new 48-multiplication algorithm for matrix operations, surpassing records from 1969.

These agents propose experimental protocols, generate code for simulations, and in some cases interface with robotic laboratory equipment. ChemCrow, for example, uses an LLM front-end with robotic tools for autonomous chemical synthesis, successfully producing 29 organosilicon compounds including 8 novel ones.

End-to-end systems generate complete manuscripts with methods, results, and discussion sections. The quality is evaluated through both automatic metrics and human review, with recent systems approaching the clarity of human-written research papers.

# Simplified EvoScientist-style agent loop
class AIScientistAgent:
    def __init__(self, llm, memory):
        self.researcher = ResearcherAgent(llm, memory.ideation)
        self.engineer = EngineerAgent(llm, memory.experimentation)
        self.evolution_manager = EvolutionManager(memory)
 
    def research_cycle(self, topic):
        # Generate and rank ideas using persistent memory
        ideas = self.researcher.generate_ideas(topic)
        ranked = self.researcher.rank_by_novelty_feasibility(ideas)
 
        for idea in ranked[:3]:
            # Implement and run experiments
            code = self.engineer.implement(idea)
            results = self.engineer.execute(code)
 
            # Evolve memory with new insights
            self.evolution_manager.update(idea, results)
 
        return self.researcher.synthesize_paper(ranked, results)

• Validation: Autonomous results require human verification for correctness
• Reproducibility: Ensuring experiments can be independently replicated
• Scalability: Extending beyond narrow domains to broader scientific inquiry
• Creativity: Generating truly novel insights vs. recombining existing knowledge
• Safety: Preventing harmful experimental designs in wet-lab settings

• EvoScientist: Multi-Agent Evolving AI Scientists for End-to-End Scientific Discovery (arXiv:2603.08127)
• Kosmos: Multimodal Scientific Discovery (arXiv:2511.02824)
• InternAgent-1.5: Unified Framework for Long-Horizon Scientific Discovery (arXiv:2602.08990)
• Sakana AI Scientist
• Frontiers: AI Agents for Scientific Discovery

• Continual Learning Agents
• Long-Horizon Agents
• Agent Trajectory Optimization