Part IV: Discovery Through Knowledge
Chapter 40: Research Agents

Research Agents

"I gave the literature agent a broad query and came back to find it had cited 200 papers, written a review article, and submitted it to arXiv under its own name."

An Agent That Forgot Its System Prompt

Overview

The previous chapters in this part built individual capabilities: mining literature (Chapter 36), retrieving relevant context (Chapter 37), constructing knowledge graphs (Chapter 38), and generating hypotheses (Chapter 39). A researcher uses all of these capabilities together, switching fluidly between reading, coding, analyzing, and reviewing. Research agents are AI systems that orchestrate these capabilities into coherent scientific workflows.

This chapter introduces four classes of research agents (literature, coding, analysis, and reviewer), examines landmark systems that combine them (Coscientist, ChemCrow, Google's AI Co-Scientist), and teaches you to build your own multi-agent research team. The agent memory systems you learn here (episodic, semantic, procedural) give these agents the ability to accumulate knowledge across sessions, while optimal stopping theory tells them when to stop searching and start acting.

Where Chapter 17 built multi-agent teams for software engineering, this chapter applies the same coordination patterns to scientific research. The roles change (from developer and tester to experimentalist and reviewer), but the underlying architecture of typed handoffs, debate loops, and human gates carries over directly. The systems built here are the stepping stones to the fully autonomous AI scientists of Chapter 53.

Prerequisites

You should have read Chapter 36: Literature Mining for text extraction and semantic search over papers, Chapter 37: Retrieval Augmented Discovery Systems for the RAG pipeline, and Chapter 39: Hypothesis Generation for how agents propose testable claims. Familiarity with multi-agent workflow patterns from Chapter 17 (agent roles, workflow graphs, debate) is essential for Section 40.3. Basic tool-calling mechanics from Chapter 10 are assumed throughout.

Learning Outcomes

Sections

40.1 Literature, Coding, and Analysis Agents

The four canonical research agent roles: literature agents that search, retrieve, and synthesize papers; coding agents that write experiment scripts and data pipelines; analysis agents that run statistical tests and interpret results; and reviewer agents that critique methodology and check for errors. Agent memory architectures: episodic memory for session history, semantic memory for domain knowledge, and procedural memory for learned protocols. Optimal stopping theory for deciding when to stop searching and start acting.

40.2 AI Scientists and Coscientist

Landmark research agent systems: Coscientist (Nature 2023) for autonomous chemical synthesis planning, ChemCrow for chemistry tool use, and Google AI Co-Scientist for hypothesis generation through debate and ranking. Architectural patterns that distinguish each system. The tournament selection mechanism in AI Co-Scientist. Safety considerations and the role of human oversight in autonomous research.

40.3 Building a Research Agent Team

Recipe: a multi-agent research team modeled on AI Co-Scientist with debate and ranking. Implementation with OpenAI Agents SDK and LangGraph. Tools integration: PaperQA2 for literature, DSPy for structured pipelines, MLflow for experiment tracking. Testing, cost control, and evaluation metrics for research agent systems. Integration with the Discovery Workbench.

What's Next

Research agents generate hypotheses, run experiments, and produce findings. But how do you know if those findings are true? Chapter 41: Scientific Claim Validation addresses the critical downstream problem: verifying that claims produced by research agents (or by human scientists) are supported by evidence, reproducible, and statistically sound. The reviewer agent from this chapter becomes one input to the validation pipeline; Chapter 41 adds systematic evidence checking, contradiction detection, and reproducibility assessment.

Bibliography

Landmark Research Agent Systems

Boiko, D. A., MacKnight, R., Kline, B., & Gomes, G. (2023). Autonomous chemical research with large language models. Nature, 624, 570-578.

Introduces Coscientist, the first LLM-driven system to autonomously plan and execute chemical experiments including Suzuki and Sonogashira cross-coupling reactions. Demonstrates that GPT-4 with tool access can reason about chemical procedures, search documentation, and control laboratory hardware.

Bran, A. M., Cox, S., Schilter, O., Baldassari, C., White, A. D., & Schwaller, P. (2024). Augmenting large language models with chemistry tools. Nature Machine Intelligence, 6, 525-535.

Presents ChemCrow, an LLM agent augmented with 18 expert chemistry tools for tasks including synthesis planning, safety checking, and molecular property prediction. Demonstrates that domain-specific tool augmentation outperforms bare LLM reasoning on chemistry tasks.

Gottweis, J., et al. (2025). AI Co-Scientist. Google DeepMind Technical Report.

Describes a multi-agent system for scientific hypothesis generation that uses a generate-debate-rank loop inspired by evolutionary computation. Validated across drug discovery, antimicrobial resistance, and rare disease research with domain expert evaluation.

Agent Architectures and Memory

Wang, L., et al. (2024). A Survey on Large Language Model based Autonomous Agents. Frontiers of Computer Science, 18(6).

Comprehensive survey of LLM agent architectures, covering memory systems (episodic, semantic, procedural), planning strategies, and multi-agent interaction patterns. Provides the taxonomy used in Section 40.1.

Park, J. S., et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023.

Introduces the memory stream architecture where agents store experiences as natural language records, retrieve them by recency, importance, and relevance, and reflect on them to form higher-level abstractions. The foundation for episodic memory in research agents.

Lála, J., et al. (2023). PaperQA: Retrieval-Augmented Generative Agent for Scientific Research. arXiv:2312.07559.

Presents PaperQA2, an agent that retrieves, reads, and synthesizes information from scientific papers with full citation support. Achieves superhuman performance on literature review tasks as measured by expert evaluation.

Tools and Frameworks

OpenAI. (2025). OpenAI Agents SDK.

Lightweight Python framework for building multi-agent workflows with handoffs, guardrails, and tracing. Used in Section 40.3 for the research team implementation.

LangChain. (2024). LangGraph: Build Stateful Multi-Actor Applications.

Graph-based framework for stateful agent workflows with typed state, conditional edges, persistence, and streaming. Used in Section 40.3 for the debate-and-rank loop.

Khattab, O., et al. (2024). DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines. ICLR 2024.

A framework for programming LLM pipelines declaratively with automatic prompt optimization. Used in Section 40.3 for structured research agent modules.

MLflow. (2024). MLflow: An Open Source Platform for the ML Lifecycle.

Experiment tracking, model registry, and deployment platform. Used in Section 40.3 for logging multi-agent research runs, tracking costs, and comparing agent configurations.

Optimal Stopping and Decision Theory

Ferguson, T. S. (2006). Optimal Stopping and Applications. UCLA Mathematics.

The standard reference on optimal stopping theory, including the secretary problem and its generalizations. Provides the mathematical foundation for the search-vs-act decision in Section 40.1.

Du, Y., et al. (2023). Improving Factuality and Reasoning in Language Models through Multiagent Debate. arXiv:2305.17126.

Shows that multi-agent debate improves factual accuracy and reasoning performance. The theoretical basis for the debate mechanism in AI Co-Scientist and in our recipe in Section 40.3.