Factory for the Plan-and-Execute agent pattern. More...

#include <neograph/graph/engine.h>

Include dependency graph for plan_execute_graph.h:

Functions
std::unique_ptr< GraphEngine >	neograph::graph::create_plan_execute_graph (std::shared_ptr< Provider > provider, std::vector< std::unique_ptr< Tool > > tools, const std::string &planner_prompt, const std::string &executor_prompt, const std::string &responder_prompt, const std::string &model="", int max_step_iterations=5)
	Build a Plan-and-Execute graph.

Detailed Description

Factory for the Plan-and-Execute agent pattern.

Produces a graph with three phases:

start → planner → [plan_empty? responder : executor] executor → [plan_empty? responder : executor] responder → end

Phase roles:

planner: Calls the LLM with planner_prompt. The model must reply with a JSON array of strings (one step per item); the node tolerates fenced ``json blocks and leading prose. The array is written to theplanchannel.
executor: Pops the next step offplan, runs an inner ReAct loop against that single step using @p executor_prompt and the shared tool set, and appends {step, result} to the past_stepschannel.
responder: Composes a final answer from the original objective and the accumulatedpast_steps, using @p responder_prompt. The text lands on thefinal_responsechannel and is also appended tomessages` as an assistant turn.

The factory registers its three custom node types and the plan_empty condition the first time it is called (idempotent via std::call_once).

Function Documentation

std::unique_ptr< GraphEngine > neograph::graph::create_plan_execute_graph	(	std::shared_ptr< Provider >	provider,
		std::vector< std::unique_ptr< Tool > >	tools,
		const std::string &	planner_prompt,
		const std::string &	executor_prompt,
		const std::string &	responder_prompt,
		const std::string &	model = `""`,
		int	max_step_iterations = `5`
	)

Build a Plan-and-Execute graph.

Parameters

provider	LLM provider shared by every phase.
tools	Tools the executor may invoke (ownership transferred).
planner_prompt	System prompt for the planner phase. Should instruct the model to respond with a JSON array of steps.
executor_prompt	System prompt for the single-step executor.
responder_prompt	System prompt for the final synthesis phase.
model	Model name. Empty falls back to the provider default.
max_step_iterations	Max tool-call iterations inside the executor per step (safety bound on the inner ReAct loop).