AI Agent
An AI system that takes sequences of actions autonomously — calling tools, browsing the web, writing and running code — to complete a multi-step goal without constant human oversight at each step.
What it means
An AI agent is a system that combines a language model with the ability to take actions — executing code, calling APIs, searching the web, reading and writing files — in an autonomous loop. The model observes the result of each action, decides on the next step, and continues until it reaches a goal or determines it cannot proceed.
The key property that distinguishes an agent from a simple chatbot is the ability to take actions that affect the world, not just produce text. An agent framework typically provides:
- Tools the model can invoke (web search, code execution, file access, API calls)
- A loop that feeds tool outputs back to the model as context for the next decision
- State management to track what has been done across multiple steps
Common frameworks for building research agents include LangChain, LlamaIndex, and Anthropic’s own Claude tool-use API. Tools like Elicit use agentic pipelines internally to search and extract data across many papers.
Why it matters for researchers
Agents automate workflows that would otherwise require manual orchestration. A literature review agent might: search PubMed for relevant papers, download and parse PDFs, extract key data from each, and return a structured summary — all without the researcher managing each step.
The autonomy-risk tradeoff: The more steps an agent takes autonomously, the more opportunities for errors to compound undetected. A single wrong API call early in a pipeline can cascade into downstream failures that look plausible but are incorrect. For research, this makes human-in-the-loop checkpoints important for high-stakes tasks.
Agentic use in research tools today:
- Elicit’s Research Agents run automated search and extraction workflows across large paper sets
- Coding assistants like GitHub Copilot Workspace plan and implement multi-file code changes
- Notebook environments with code execution allow models to run analyses and iterate on results
What to watch for: When a tool says it “automatically” extracted data from hundreds of papers, ask what the agent is doing internally — what tools it called, what prompts it used, and how errors are handled. This affects the reproducibility of any results built on agentic pipelines.