Agentic AI Tools

22.2. Agentic AI Tools#

Agentic tools read code, write and run it, call other tools, check the results, and repeat until a task is done. They differ in how much autonomy they take, from suggesting an edit you approve to running a whole task on their own. This page lists widely used tools for research; to run them on the cluster, see Using Agentic AI on the Cluster.

Note

This area moves fast, and the lists are representative, not complete. Check a tool’s current status and pricing before you adopt it. To suggest a tool or a change, open an issue in the computing handbook GitHub repository.

22.2.1. Coding assistants and agents#

The most common agentic tools write and edit code across a whole project. Most run as a terminal command, a VS Code extension, or both.

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
opencode	Open-source terminal coding agent that works with many model providers, including your own endpoint
Goose	Open-source agent, as a CLI and a desktop app, now part of the Agentic AI Foundation at the Linux Foundation
Qwen Code	Open-source terminal agent, originally based on Gemini CLI, for Qwen and other models through OpenAI, Anthropic, and Gemini APIs

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

Cloud coding agents run in the provider’s cloud rather than on your machine. You give one a task on a GitHub repository; it works in a temporary cloud environment and opens a pull request for you to review.

Tool	What it is
Claude Code on the web	Claude Code sessions on Anthropic-managed infrastructure, started from claude.ai/code or with `claude --cloud`
OpenAI Codex cloud	Codex working through coding tasks in the cloud, then committing or opening a pull request
GitHub Copilot cloud agent	Copilot researching a repository, planning, and implementing changes in an ephemeral cloud environment
Jules	Google’s experimental coding agent, which works in a cloud virtual machine and opens a pull request

Claude Code can also run inside GitHub Actions, responding when someone mentions @claude in an issue or pull request.

Warning

A cloud coding agent copies your repository to the provider’s infrastructure, so the code, and any data or secrets committed to it, leave Harvard. Use these agents only on repositories whose contents are cleared for that service, keep data and credentials out of the repository, and review each pull request as you would an outside contribution. See Security and Compliance.

22.2.2. Research and science agents#

Beyond coding, agentic tools can take on parts of the research process: searching 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
Ai2 Asta	Research assistant from the Allen Institute for AI for finding papers, summarizing literature, and analyzing data
Semantic Scholar	Free academic search engine from the Allen Institute for AI, covering more than 200 million papers

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, Claude Research, and Perplexity. For a workflow that keeps every citation checkable, see Literature Review and Data Exploration.

22.2.3. 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

22.2.4. Model Context Protocol (MCP)#

The Model Context Protocol (MCP) is an open standard that connects agents to external tools, data, and services, such as files, databases, or APIs. Most of the tools above are MCP clients, so a server you write or install works with all of them; see Building Custom Tools and MCP Servers.

22.2.5. Choosing a tool#

Before you adopt 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. Then consider:

Cloud or local. Cloud models (Claude, GPT, Gemini) are often the most capable, but they send your prompts and code to an external provider. To keep everything on the cluster, serve an open-weight model yourself; see HPC Agentic Recipes. Open-source agents such as opencode, Goose, Qwen Code, Cline, and Aider can connect to any OpenAI-compatible endpoint, including one on the cluster.
Data sensitivity. Do not enter confidential data (Level 2 and above) into a public AI service. Harvard’s approved tools have data-protection agreements, each for a specific data level; for example, the Harvard AI Sandbox is approved up to Level 3. Check a tool’s approved level before you use it with the cluster’s data; on the cluster, the rule in Before you start also applies. 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. Terminal agents run in an SSH session, and IDE extensions connect through VS Code Remote-SSH (see VSCode for Remote Dev). Both work on the cluster.