Hybrid agent

Holo4 is trained on tasks that cross interfaces, so it does not need a separate agent per tool. This cookbook starts from the browser agent and adds the Python runner from the code agent. Asked to find two dates and compute the gap between them, Holo reads the dates off the page and leaves the arithmetic to Python.

This example runs model-written code on your machine with subprocess. Outside a demo, run it in a container or a sandbox with no network and no secrets.

Start from the browser agent

Install and save browser_agent.py, then make the edits below. Nothing else changes.

Add the runner

Import subprocess and sys, and register run_python next to the screen tools:

browser_agent.py

import base64
import json
import os
import subprocess
import sys

# ...

tools = [
    fn("click", "Click at (x, y) coordinates", element={"type": "string", "description": "Detailed description of the target UI element"}, x=INT, y=INT),
    fn("type", "Type text into the focused element, optionally pressing Enter", text={"type": "string"}, press_enter={"type": "boolean"}),
    fn("scroll", "Scroll the page", direction={"type": "string", "enum": ["up", "down"]}),
    fn("goto", "Navigate to a URL", url={"type": "string"}),
    fn("run_python", "Run a Python script and return its stdout and stderr. Print anything you want to see.", code={"type": "string", "description": "Python source code"}),
    fn("answer", "Provide a final answer", content={"type": "string", "description": "The answer content"}),
]

Route the call

Handle run_python first in execute, before the screen actions:

browser_agent.py

def execute(page, name: str, args: dict) -> str:
    if name == "run_python":
        try:
            proc = subprocess.run([sys.executable, "-c", args["code"]], capture_output=True, text=True, timeout=30)
        except subprocess.TimeoutExpired:
            return "Timed out after 30 seconds"
        return (proc.stdout + proc.stderr)[-4000:] or "(no output)"
    if name == "click":
        # ... unchanged

Update the prompt and task

browser_agent.py

{"role": "system", "content": "You are a web agent. You see the browser through screenshots and act with your tools. Use run_python for any computation."},

# ...

print(run("On Wikipedia, find Ada Lovelace's birth and death dates, then compute exactly how many days she lived.", "https://en.wikipedia.org"))

Run it

python browser_agent.py

Output

--- click({'element': 'Search bar at the top of Wikipedia', 'x': 359, 'y': 39})
--- type({'text': 'Ada Lovelace', 'press_enter': True})
--- run_python({'code': 'from datetime import date\n\nbirth = date(1815, 12, 10)\ndeath = date(1852, 11, 27)\ndays_lived = (death - birth).days\nprint(f"Days lived: {days_lived}")'})
Ada Lovelace was born on December 10, 1815 and died on November 27, 1852. She lived for exactly 13,502 days.

Two screen steps to read the dates, one code step to count the days.

How it works

One loop, one toolbox. The model sees every tool on every step and chooses by what the step needs. There is no router and no second agent.
Screenshots keep coming. Every step still sends a fresh screenshot, including after run_python. The page has not changed, and the model keeps its view of it.
Add more interfaces the same way. MCP tools from the MCP agent slot into the same tools list; route their names to session.call_tool.

Get started

Build an agent

Cookbooks

Guides

Self-host

Start from the browser agent

Add the runner

Route the call

Update the prompt and task

Run it

How it works

Next steps

Function calling

Core concepts

Start from the browser agent

Add the runner

Route the call

Update the prompt and task

Run it

​How it works

​Next steps

Function calling

Core concepts

How it works

Next steps