jeevesagent.architecture.plan_and_execute

Plan-and-Execute: planner → step executor → synthesizer.

Wang et al. 2023 (Plan-and-Solve Prompting), He et al. 2025 (Plan-then-Execute Pattern Implementation). The 2026 production pattern for cost-sensitive multi-step tasks where ReAct-style “think before each action” wastes tokens — Plan-and-Execute commits to a plan upfront and executes step-by-step.

Pattern (v1 simple)

1. Plan. A single planner call produces a JSON list of step descriptions. Steps are short, ordered, dependency-free for v1.

2. Execute. Each step is a text-only model call seeded with the original problem, the plan, and prior step outputs. Sequential.

3. Synthesize. A final model call combines step outputs into the answer.

What’s NOT in v1

• Sub-architecture invocation per step. Each step is a single text-only call, NOT a fresh ReAct (or any other architecture) invocation. That’s the planned v0.5 work — it unblocks Deep Agent and a richer ReWOO too.

• Replanning on failure. max_replans is in the constructor but not yet wired; it’ll land alongside the sub-architecture invocation primitive when steps can fail meaningfully.

• Parallel step execution. Steps run sequentially. DAG-aware parallel execution is straightforward to add once dependencies are first-class on PlanStep.

Strengths

• Cheaper than ReAct for tasks with predictable structure: one planner + N step calls + one synthesizer ≪ N×K ReAct turns.

• Observable plan. The plan is a structured Pydantic object you can log, audit, or override before execution.

Weaknesses

• No tool use within steps in v1 (use ReAct or Supervisor for tool-heavy work).

• Plan quality bounds answer quality; bad planner = bad output.

Attributes

DEFAULT_EXECUTOR_PROMPT
DEFAULT_PLANNER_PROMPT
DEFAULT_SYNTHESIZER_PROMPT

Classes

Plan	A list of plan steps in execution order.
PlanAndExecute	Planner → step executor → synthesizer.
PlanStep	One step of a plan.
StepResult	The output of executing one step.

Module Contents

class jeevesagent.architecture.plan_and_execute.Plan(/, **data: Any)

Bases: pydantic.BaseModel

A list of plan steps in execution order.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

steps: list[PlanStep] = None

class jeevesagent.architecture.plan_and_execute.PlanAndExecute(*, max_steps: int = 8, planner_prompt: str | None = None, executor_prompt: str | None = None, synthesizer_prompt: str | None = None)

Planner → step executor → synthesizer.

declared_workers() → dict[str, jeevesagent.agent.api.Agent]

async run(session: jeevesagent.architecture.base.AgentSession, deps: jeevesagent.architecture.base.Dependencies, prompt: str) → collections.abc.AsyncIterator[jeevesagent.core.types.Event]

name = 'plan-and-execute'

class jeevesagent.architecture.plan_and_execute.PlanStep(/, **data: Any)

Bases: pydantic.BaseModel

One step of a plan.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

description: str

id: str

class jeevesagent.architecture.plan_and_execute.StepResult(/, **data: Any)

Bases: pydantic.BaseModel

The output of executing one step.

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

description: str

output: str

step_id: str

jeevesagent.architecture.plan_and_execute.DEFAULT_EXECUTOR_PROMPT = Multiline-String

Show Value

"""You are executing one step of a multi-step plan. Use the original
task and the prior step outputs to produce this step's output.

Output ONLY the result of this step (no preamble, no commentary)."""

jeevesagent.architecture.plan_and_execute.DEFAULT_PLANNER_PROMPT = Multiline-String

Show Value

"""You produce a step-by-step plan to solve the user's task. The plan
must be executable: each step should be a concrete sub-task that an
LLM can complete in a single response.

Output a JSON list of strings — each string is one step description.
Aim for 3-7 steps; fewer is better when the task is simple. Order
matters: each step can use prior step outputs.

Output ONLY the JSON list. No prose, no markdown fences."""

jeevesagent.architecture.plan_and_execute.DEFAULT_SYNTHESIZER_PROMPT = Multiline-String

Show Value

"""You synthesize the final answer from a sequence of step outputs.
Combine them into a single coherent response that addresses the
original task. Be concise; don't repeat the steps verbatim."""