Skip to main content
Version: v1.7.3-10 🚧

Agents

What's the H2O AI super agent?​

Enterprise h2oGPTe agents are general-purpose, code-first AI agents. They pick tools, write code, run it, fix errors, and keep iterating until they finish the task.

Unlike frameworks such as LangGraph or CrewAI, where each agent targets one specific workflow, an H2O agent isn't tied to a playbook. The same request can produce different solutions on different runs because the agent adapts to context, available tools, and what it learned from earlier steps.

The Agents button in the chat composer switches a chat session from question-and-answer to autonomous task execution.

H2O AI Super Agent overview

Proven performance: #1 on FutureX​

The H2O AI Super Agent ranks #1 on the FutureX leaderboard, a contamination-free benchmark from researchers at ByteDance Seed, Fudan University, Stanford University, and Princeton University that tests AI agents on predicting real-world future events. H2O.ai holds three of the top four positions, ahead of agents from OpenAI, Google, DeepSeek, and xAI.

FutureX evaluates around 500 events per week across 11 domains (finance, technology, sports, politics, healthcare, and more) from 195 websites, with difficulty-weighted scoring.

What agents can do out of the box​

With no configuration, an agent has access to:

Data and AI
Run AutoML experiments, analyze datasets, generate charts, and build ML models with Python.
Web and research
Search across Google, Bing, Wikipedia, academic journals, and internal sources.
Content and media
Generate images, build presentations, transcribe audio and video, and capture page screenshots.
Code and automation
Write, test, and debug code. Run shell scripts. Integrate with GitHub.
Documents
Ask questions about uploaded files, extract insights, summarize content, and auto-tag by type.
External services
Connect to third-party APIs and databases through MCP servers.

See Agent tool configuration for the full list of built-in tools and their configuration parameters.

Agent flavors​

The General Agent is the core H2O AI Super Agent. With the right settings and prompt, it can handle most tasks. h2oGPTe also includes purpose-built flavors for common workflows, plus specialized flavors for building tools and custom agents.

Core flavors​

General Agent
Broadest tool access, including code execution and the browser. Best when the task could go in many directions.
Deep Research
A General Agent variant with self-critique. Use for in-depth investigation that benefits from review loops.
Data Science Agent
Data augmentation, exploratory data analysis, and model building, with H2O Driverless AI integration.
Code Agent
Code generation, debugging, reviews, and software development workflows.

Specialized flavors​

Deep Research Zero
Builds its own tools from scratch with a minimal base toolset. Use when the task needs tools the platform doesn't ship with.
Web search
Targeted web searches when you don't need code execution or other tools.
Tool Builder
Build MCP tools, browser actions, or general tools that other agents can use.
Agents Builder
Generate custom agents in LangGraph, OpenAI Agents SDK, or CrewAI from a natural-language description.
MCP-only Runner
Run tasks using only MCP tools, without other built-in capabilities.

To select an agent flavor:

  1. From the h2oGPTe dashboard, click + New chat.
  2. Click the Agents button in the chat composer.
  3. Pick the flavor that matches your task.
tip

Start with General Agent if you're not sure. The specialized flavors just get to the answer faster on their target task.

Accuracy levels​

Each flavor supports four accuracy levels. They control how many turns the agent can take and how long it can spend per turn.

LevelTurnsTimeUse it for
Quick1030 sFast lookups and simple answers
Basic2060 sSingle-step tasks
Standard40120 sMost tasks (default)
Maximum80240 sComplex analysis where accuracy matters more than speed

Higher accuracy improves results on hard tasks at the cost of time.

Customize agents with collections​

Connect an agent to a collection to give it access to your documents. The agent can search and reason over the collection alongside its built-in tools, which turns a general-purpose agent into a domain-specific one.

RAG modes​

How the agent interacts with your documents depends on the RAG (retrieval-augmented generation) mode:

ModeHow it worksBest for
Agentic RAGThe agent decides when to query the documentsComplex questions that combine document knowledge with other tools
RLM RAGA retrieval-language model with up to two retrieval iterationsFocused document questions where you want concise, grounded answers
Fast Agentic RAGDocument chunks load directly into the agent's contextSpeed-sensitive queries over smaller document sets

Custom tools​

Upload your own code as a tool. A custom tool can:

  • Accept structured inputs (for example, a company name).
  • Run custom business logic (for example, look up that company's brand colors).
  • Return a result the agent can use in the next step (for example, apply those colors to a chart).

Custom tools are user-owned and can be associated with agents. You configure them in the prompt settings. See Tool calling for details.

You can also turn agent-generated .py or .zip files into tools directly from the agent response, without downloading and re-uploading them. See Add files as tools.

Custom agents​

If you need a specific orchestration pattern, build a custom agent in one of the supported frameworks:

  • LangGraph for graph-based workflows.
  • CrewAI for multi-agent collaboration.
  • OpenAI Agents SDK for OpenAI-compatible agents.
  • Claude for Anthropic Claude-based agents.

Package the agent as a ZIP file and upload it from the Agents tab.

To skip the framework work, describe what you want in the Agents Builder and h2oGPTe generates the code. The Agents Builder produces LangGraph, OpenAI Agents SDK, or CrewAI code. For Claude agents, package and upload the ZIP yourself.

How the pieces connect​

An h2oGPTe agent brings together these parts of the platform:

  • Models: Agents use the LLM configured for the chat session (Claude, GPT, a local model, or a custom one). The model handles reasoning and tool selection.
  • Tools: Built-in tools cover core capabilities. Custom tools and MCP servers extend the agent with your own code and external integrations.
  • Files: Agents read uploaded files and can generate new ones (scripts, charts, documents, presentations) that you can download.
  • Collections and RAG: Connect agents to document collections for domain-specific knowledge. Pick a RAG mode that fits your accuracy and speed needs.
  • Access control: Admins control which tools and models agents can access, and who can create or manage custom agents and API keys.
  • Evaluation: Use Eval Studio to score agent responses, compare configurations, and track performance over time.

Review agent behavior​

To see how an agent completed (or is completing) a request:

  1. Choose a view mode. Click View at the top-right of the agent response to switch between Simplified (final response only), Standard (chain of thoughts visible), or Detailed (chain of thoughts plus the right panel opens automatically). See Agent response view for details.
  2. Open the right panel. Click Details to open the right panel. It contains tabs for Steps (an organized timeline of code generation, execution, and results), Files (workspace file explorer), Sources (external web references), Thoughts (how the agent chooses tools and plans its approach), and Chat (internal agent-to-agent logs).
  3. Review the plan before execution. When the agent has a plan ready, review it. Approve it as-is or change it before the agent runs.
  4. Download outputs. In the right panel, click Files, then click the menu icon. From there, you can Download files, Import to collection, Add files as tools, or — if your admin has enabled custom agents — Add Files As Agent.

Tutorials​

Feedback