Skip to main content
Agent tools let one agent delegate tasks to another. An orchestrator agent calls specialized sub-agents as tools—each with its own environment, model, and evaluation.

​
Why Sub-Agents?

Complex tasks often break into specialized subtasks:
  • Orchestrator decides what to do
  • Researcher gathers information
  • Coder writes code
  • Reviewer checks work
Instead of one massive agent trying to do everything, you compose focused specialists. 
Orchestrator
    ├── call research_tool("find security issues")
    │       └── ResearchAgent runs with browsing tools
    ├── call code_tool("fix the vulnerability")
    │       └── CodingAgent runs with bash/edit tools
    └── call review_tool("verify the fix")
            └── ReviewAgent runs read-only

​
AgentTool

Wraps a Task template so it can be called as a tool. 
from hud import Environment
from hud.tools import AgentTool

# Define a specialist environment
researcher_env = Environment("researcher")
researcher_env.add_tool(PlaywrightTool())
researcher_env.add_tool(WebSearchTool())

@researcher_env.scenario()
async def investigate(issue_id: str):
    response = yield f"Research issue {issue_id} and summarize findings"
    yield 1.0  # Evaluation based on response

# Create tool from task template
research_tool = AgentTool(
    researcher_env("investigate"),
    model="gpt-4o",
    name="research",
    description="Research an issue and return findings",
)
Now any agent can call research(issue_id="SEC-123") and a full sub-agent runs.

​
Parameters

AgentTool(
    task,                          # Task template to run
    model="gpt-4o",                # Model for the sub-agent
    # OR
    agent=CustomAgent,             # Custom agent class
    agent_params={"max_steps": 5}, # Passed to agent.create()
    name="research",               # Tool name
    description="...",             # Tool description
    trace=True,                    # Enable tracing for debugging
)
Use model= for standard agents or agent= for custom agent classes.

​
Eval-Only Parameters

Parameters with | None = None are hidden from the tool schema but available for evaluation: 
@env.scenario()
async def investigate(
    issue_id: str,                      # Orchestrator sees this
    expected_cause: str | None = None,  # Hidden - for eval only
):
    response = yield f"Research issue {issue_id}"
    
    # Use expected_cause in evaluation
    if expected_cause and expected_cause in response:
        yield 1.0
    else:
        yield 0.5
The orchestrator calls investigate(issue_id="SEC-123"). The eval harness can pass expected_cause for automated scoring.

​
Orchestrator Setup

from hud import Environment
from hud.tools import AgentTool, BashTool, EditTool
from hud.agents import create_agent
import hud

# Specialist environments
researcher_env = Environment("researcher")
# ... setup researcher tools and scenario

coder_env = Environment("coder")
coder_env.add_tool(BashTool())
coder_env.add_tool(EditTool())

@coder_env.scenario()
async def fix_bug(description: str):
    yield f"Fix the bug: {description}"
    yield 1.0

# Orchestrator with sub-agent tools
orchestrator = Environment("orchestrator")
orchestrator.add_tool(AgentTool(
    researcher_env("investigate"),
    model="gpt-4o",
    name="research",
))
orchestrator.add_tool(AgentTool(
    coder_env("fix_bug"),
    model="claude-sonnet-4-5",
    name="fix_code",
))

@orchestrator.scenario()
async def handle_ticket(ticket_id: str):
    yield f"Handle support ticket {ticket_id}"
    yield 1.0

# Run orchestrator
task = orchestrator("handle_ticket", ticket_id="TICKET-456")
agent = create_agent("gpt-4o")

async with hud.eval(task) as ctx:
    await agent.run(ctx)

​
Tracing

Enable trace=True to record sub-agent execution: 
research_tool = AgentTool(
    researcher_env("investigate"),
    model="gpt-4o",
    trace=True,  # Records full sub-agent trajectory
)
Sub-agent traces are linked to the parent trace for hierarchical debugging.

​
When to Use

Good for:
  • Complex workflows with distinct phases
  • Mixing models (fast for orchestration, powerful for coding)
  • Isolating tool access (researcher can’t edit files)
  • Reusable specialist agents
Avoid when:
  • Simple linear tasks
  • Latency-sensitive applications (sub-agent overhead)
  • Single-model workflows

​
Tips

Keep sub-agents focused. One clear task per specialist. Match models to complexity. Use cheaper models for simple delegation, expensive ones for hard problems. Test specialists independently. Run each sub-agent scenario directly before composing. → Computer Tools — GUI automation for sub-agents → Coding Tools — Shell and editing for coding agents