AI / 大模型架构

多智能体框架

由有向图组织的多智能体协同评估工作流，含任务分解、知识检索、推理与校验。

完整 Prompt

A multi-agent collaboration framework with four specialised agents arranged in a closed loop.

Center-top: User Query (rounded rectangle).

Four agents arranged clockwise around a central message bus:
1) Planner — decomposes the user task into sub-goals and dispatches them.
2) Retriever — issues semantic queries against an external Vector Store + Document Store (drawn outside the loop on the left).
3) Executor — calls external Tools / APIs (drawn outside the loop on the right) to perform actions.
4) Critic — evaluates intermediate outputs against the user goal and returns revisions.

Show labeled directional arrows for each message type:
- "task" (Planner -> Retriever / Executor)
- "context" (Retriever -> Planner)
- "action result" (Executor -> Critic)
- "revision" (Critic -> Planner, dashed)

Bottom: Final Response box.

Style: clean academic vector, navy / teal / amber palette, rounded boxes, white background, sans-serif labels. Suitable for ICLR or AAAI submissions.

立即试用此 Prompt

适用场景

适合智能体协同、任务分解、自动化评估系统的架构图（中文论文）。

变体

Two-agent reviewer-writer loop

A two-agent loop: Writer agent drafts a long-form answer; Reviewer agent scores the draft on 4 criteria (correctness, completeness, conciseness, citation coverage) and returns structured feedback. The loop terminates when score crosses a threshold T. Show the score-gated termination as a diamond decision node. Clean academic style.

Hierarchical orchestrator

A hierarchical agent system: a top-level Orchestrator agent dispatches to 3 specialist sub-agents (Code, Search, Math), each of which manages 2 worker tools. Show the 3-level tree clearly with labeled message types. Use a clean tree layout with rounded boxes.

使用建议

"思维图" 比 "图调度" 更具体，模型会画出有向图结构。
消息类型标在箭头上是必要的，否则箭头空白难读。
反馈弧线让"持续优化"立刻可见。

常见问题

How do I show parallel agent execution?

Add "Planner dispatches in parallel; show fork/join with horizontal bars" or describe a swimlane layout.

Can I add a memory module shared by all agents?

Yes — describe it as a central long-term memory store that every agent reads from and writes to with bidirectional arrows.