agents.planning.llm_compiler.agent¶

LLM Compiler Agent Implementation.

from typing import Any, Dict This implementation follows the LLM Compiler architecture from the paper by Kim et al., focusing on parallelizable task execution through a DAG structure.

Classes¶

LLMCompilerAgent

LLM Compiler Agent implementation.

Functions¶

main()

Main.

Module Contents¶

class agents.planning.llm_compiler.agent.LLMCompilerAgent(config)¶

Bases: haive.core.engine.agent.agent.AgentArchitecture

LLM Compiler Agent implementation.

This agent architecture has three main components: 1. Planner: Creates a task DAG 2. Task Executor: Executes tasks as their dependencies are satisfied 3. Joiner: Processes results and decides whether to output an answer or replan

Initialize the LLM Compiler agent.

Parameters:: config (agents.planning.llm_compiler.config.LLMCompilerAgentConfig)

async arun(query)¶

Run the agent asynchronously.

Parameters:: query (str) – The user’s query
Returns:: Response from the agent

execute_tasks(state)¶

Execute tasks in parallel as their dependencies are satisfied.

Parameters:: state (agents.planning.llm_compiler.state.CompilerState) – Current agent state
Returns:: Updated state with executed task results
Return type:: dict[str, Any]

join(state)¶

Process the results and decide whether to provide a final answer or replan.

Parameters:: state (agents.planning.llm_compiler.state.CompilerState) – Current agent state
Returns:: Decision to end or replan
Return type:: dict[str, Any]

plan(state)¶

Generate a plan based on the user’s query.

Parameters:: state (agents.planning.llm_compiler.state.CompilerState) – Current agent state
Returns:: Updated state with a new plan
Return type:: dict[str, Any]

run(query)¶

Run the agent on a query.

Parameters:: query (str) – The user’s query
Returns:: Response from the agent

setup_workflow()¶

Set up the agent workflow as a state graph.

Return type:: Any

should_execute_more(state, config=None)¶

Determine the next execution step.

Parameters:

state (CompilerState) – The current agent state.
config (Optional[Any]) – Execution configuration (not used but required).

Returns:

The next node to execute in the state graph.

Return type:

str

stream(query)¶

Stream the agent’s execution.

Parameters:: query (str) – The user’s query
Yields:: Execution steps

agents.planning.llm_compiler.agent.main()¶

Main.

Returns:: Add return description]
Return type:: [TODO