haive.agents.supervisor.agent¶

Dynamic Supervisor V2 - Main agent implementation.

This module contains the core DynamicSupervisor class that orchestrates runtime agent discovery, creation, and task routing.

Classes¶

DynamicSupervisor

Advanced supervisor that discovers and creates agents at runtime.

Functions¶

create_dynamic_supervisor([name, agent_specs, ...])

Factory function to create a configured dynamic supervisor.

Module Contents¶

class haive.agents.supervisor.agent.DynamicSupervisor(**data)¶

Bases: haive.agents.react.agent.ReactAgent

Advanced supervisor that discovers and creates agents at runtime.

The DynamicSupervisor extends ReactAgent to provide intelligent task routing with the ability to discover, create, and manage specialized agents dynamically based on task requirements.

Key capabilities:

  • Dynamic agent discovery from specifications

  • Runtime agent creation and lifecycle management

  • Intelligent task-to-agent matching

  • Performance tracking and optimization

  • Extensible discovery mechanisms

name¶

Supervisor identifier

engine¶

LLM configuration for supervisor reasoning

agent_specs¶

Initial specifications for creatable agents

discovery_config¶

Configuration for agent discovery

max_agents¶

Maximum number of active agents to maintain

auto_discover¶

Whether to automatically discover new agents

include_management_tools¶

Whether to include agent management tools

Examples

Basic usage with predefined agent specs:

supervisor = DynamicSupervisor(
    name="task_router",
    agent_specs=[
        AgentSpec(
            name="researcher",
            agent_type="ReactAgent",
            description="Research and analysis expert",
            specialties=["research", "analysis"],
            tools=[web_search_tool]
        ),
        AgentSpec(
            name="writer",
            agent_type="SimpleAgent",
            description="Content creation expert",
            specialties=["writing", "editing"]
        )
    ]
)

result = await supervisor.arun(
    "Research quantum computing and write a summary"
)

Initialize the dynamic supervisor.

Parameters:

**data – Configuration parameters

async arun(input_data)¶

Run the supervisor asynchronously.

Parameters:

input_data (str | dict[str, Any] | list[langchain_core.messages.BaseMessage]) – Task input (string, dict, or messages)

Returns:

Agent execution result

Return type:

Any

get_metrics()¶

Get supervisor performance metrics.

Returns:

Dictionary of metrics including agent stats

Return type:

dict[str, Any]

run(input_data)¶

Run the supervisor synchronously.

Parameters:

input_data (str | dict[str, Any] | list[langchain_core.messages.BaseMessage]) – Task input

Returns:

Agent execution result

Return type:

Any

classmethod validate_agent_specs(v)¶

Validate agent specifications have unique names.

Parameters:

v (list[haive.agents.supervisor.models.AgentSpec])

Return type:

list[haive.agents.supervisor.models.AgentSpec]

haive.agents.supervisor.agent.create_dynamic_supervisor(name='dynamic_supervisor', agent_specs=None, discovery_mode=AgentDiscoveryMode.MANUAL, **kwargs)¶

Factory function to create a configured dynamic supervisor.

Parameters:
Returns:

Configured DynamicSupervisor instance

Return type:

DynamicSupervisor

Examples

>>> supervisor = create_dynamic_supervisor(
...     name="task_router",
...     agent_specs=[math_spec, research_spec],
...     discovery_mode="manual",
...     max_agents=20
... )