Agentic AI Tools

23. Agentic AI Tools#

An agentic tool uses a large language model to plan and carry out multi-step tasks, not just answer a question: it can read a codebase, write and run code, call external tools, inspect the results, and iterate toward a goal. Tools differ in how much autonomy they take, from suggesting an edit you approve to running a whole task on their own. This page introduces widely used tools for research productivity; for how to run them on the cluster, see Using Agentic AI on the Cluster.

Note

This is a fast-moving area, and the list below is representative rather than exhaustive. Check each tool’s current status and pricing before adopting it. To suggest a new tool, or a change to something already listed, open an issue in the computing handbook GitHub repository.

23.1. Coding assistants and agents#

The most common agentic tools for research or engineering write and edit code across a whole project. Most run as a terminal (CLI) command, a VS Code extension, or both, so you can use them in whichever surface you prefer.

Tool	What it is
Claude Code	Anthropic’s coding agent, as a CLI and a VS Code extension
OpenAI Codex	OpenAI’s coding agent, as a CLI and an IDE extension
Gemini CLI	Google’s coding agent, as a CLI and a VS Code extension
GitHub Copilot	GitHub’s assistant with an agent mode, as an editor extension
Cline	Open-source autonomous coding agent for VS Code
Aider	Git-aware pair-programming agent in the terminal

Other tools are self-contained rather than add-ons to your editor:

Tool	What it is
Cursor	AI-native code editor with a project-wide agent mode
Devin Desktop	AI editor and agent platform from Cognition (formerly Windsurf)
OpenHands	Open-source platform for autonomous software-engineering agents

23.2. Research and science agents#

Beyond coding, agentic tools can take on parts of the research process itself: searching and synthesizing the literature, and running experiments.

Scientific literature:

Tool	What it is
Edison Scientific	Web and API platform of science agents for literature search and analysis, spun out of the nonprofit FutureHouse lab; FutureHouse’s PaperQA2 library is open source
Elicit	Research assistant for finding, screening, and summarizing papers
Consensus	Search engine that synthesizes findings across the research literature

Autonomous and domain-specific:

Tool	What it is
Claude Science	Anthropic’s multi-agent workbench for scientific research, with curated skills for genomics, proteomics, structural biology, and cheminformatics
Google AI co-scientist	Google’s multi-agent system for generating and refining research hypotheses
Biomni	Biomedical research agent (Stanford) with a large toolset and connected databases
Sakana AI Scientist	Publicly available framework that generates ideas, runs experiments, and drafts papers end to end

General assistants also offer a deep research mode that plans, searches the web, and returns a cited report, for example OpenAI Deep Research, Gemini Deep Research, and Perplexity.

23.3. Frameworks for building agents#

When you need a custom agent or pipeline rather than an off-the-shelf assistant, these frameworks help you define agents, give them tools, and orchestrate them.

Framework	What it is
Claude Agent SDK	Anthropic’s SDK for building agents on Claude
OpenAI Agents SDK	OpenAI’s lightweight multi-agent framework
LangGraph	Graph-based framework for stateful, multi-step agents
Microsoft AutoGen	Framework for multi-agent conversations and workflows
CrewAI	Role-based multi-agent orchestration
LlamaIndex	Data framework for building agents over your own data
Hugging Face smolagents	Minimal library for code-writing agents
NVIDIA NemoClaw	Open blueprints from NVIDIA and LangChain for building governed autonomous agents

23.4. Model Context Protocol (MCP)#

The Model Context Protocol (MCP) is an open standard that lets agents connect to external tools, data, and services, for example files, databases, or APIs, through a common interface. Most of the tools above can act as MCP clients, so an MCP server you write or install becomes available to all of them.

23.5. Choosing a tool#

Before adopting an external tool, check whether Harvard or FAS already provides a vetted one, and at what data level: see the HUIT AI Tool Comparison, AI Tools Available to the FAS, and the HUIT AI APIs and developer tools for programmatic access. A few further considerations narrow the field:

Cloud or local. Cloud services (Claude, GPT, Gemini) are often the most capable, but they send your prompts and code to an external provider. If you prefer to keep everything on the cluster, you can run an open-weight model yourself; see HPC Agentic Recipes.
Data sensitivity. Do not enter confidential data (Level 2 and above) into a public AI service. Harvard also provides approved tools with data-protection agreements, each cleared for a specific data level; for example, the Harvard AI Sandbox is approved for up to Level 3. Check a tool’s approved data level before using it with the cluster’s data, and see Security and Compliance and Harvard’s Generative AI Guidelines.
Cost. Cloud tools bill by usage; self-hosting uses your GPU allocation. Track both.
Form factor. Both terminal and IDE tools work on the cluster: terminal agents run directly in an SSH session, and IDE or extension tools connect through VS Code Remote-SSH (see VSCode for Remote Dev). Pick whichever fits your workflow.