Mental Model

Overview

The co-worker loop is easiest to hold as five layers. AI sits in Work and Evaluate; Intake, Act, and Return are Gradio input, tools, and UI output. Anthropic names from Building effective agents (PDF).

Applied workflows

• Evaluator-Optimizer. Worker produces a reply; Evaluator scores it against success criteria and either ends or loops the worker with feedback.
• Not used here (yet). Routing, Parallelization, Prompt Chaining, Orchestrator-Worker — one worker / evaluator pair in a LangGraph, not a multi-worker fan-out.

Layers

• 1 Intake. Gradio collects request + success criteria; run_superstep seeds graph state and a MemorySaver thread id.
• 2 Work. Worker rewrites its system prompt from criteria (and prior feedback), invokes gpt-4o-mini with bound tools.
• 3 Act. worker_router sends tool calls to ToolNode (Playwright, Serper, Wikipedia, REPL, sandbox files, Pushover), then back to Work.
• 4 Evaluate. Evaluator returns structured EvaluatorOutput: feedback, success_criteria_met, user_input_needed.
• 5 Return. On success or clarify need, graph ends; Gradio shows assistant reply plus evaluator feedback. Otherwise Evaluate routes back to Work.

Layer map

Layer stack for one Gradio Go. Diagram key: ✦ AI agent · second line = Anthropic workflow · purple = actor · white = service.

---
config:
  theme: default
  themeVariables:
    background: "#ffffff"
    fontSize: "14px"
---
flowchart TB
    person[Person]
    intake["1 Intake
request and criteria"]
    work["✦ 2 Work
Evaluator-Optimizer"]
    act["3 Act
ToolNode"]
    evaluate["✦ 4 Evaluate
Evaluator-Optimizer"]
    returnNode["5 Return
reply and feedback"]

    person -->|Go| intake --> work
    work -->|tool calls| act --> work
    work -->|text| evaluate
    evaluate -->|retry| work
    evaluate -->|done or clarify| returnNode --> person

    classDef agent fill:#e8f0ee,stroke:#0f3d36,stroke-width:2px,color:#0d0d0d
    classDef service fill:#fafaf8,stroke:#c8c8c2,stroke-width:1px,color:#0d0d0d
    classDef actor fill:#f3eef8,stroke:#6b4f9a,stroke-width:2px,color:#0d0d0d
    class work,evaluate agent
    class intake,act,returnNode service
    class person actor

Read layers top to bottom; drill into agent tabs for prompts and routing helpers.