Worker Agent

Overview

Layers 2–3 on the Mental Model. The Worker is the Sidekick graph’s doer. It rewrites a system prompt from success criteria and any prior evaluator feedback, invokes gpt-4o-mini with bound tools, and either routes to the tool node or to the Evaluator when the model replies in text.

Behavior

• Criteria-first prompt. Success criteria and optional rejection feedback are injected every turn.
• Tool-or-evaluate route. worker_router sends tool calls to ToolNode, otherwise to the evaluator (Evaluator-Optimizer producer side). See Building effective agents.
• Broad toolkit. Playwright browser, Serper search, Wikipedia, Python REPL, sandbox file tools, and Pushover push.
• Question or final. Instructions force either a clear user question or a finished answer — not both.
• Checkpointed. Runs under the Sidekick MemorySaver thread id from Gradio setup.

Worker flow

One worker step after START or after tools / evaluator retry. Diagram key: ✦ AI agent · second line = Anthropic workflow · purple = actor · white = service.

---
config:
  theme: default
  themeVariables:
    background: "#ffffff"
    fontSize: "14px"
---
flowchart TB
    person[Person]
    stateIn[State messages and criteria]
    worker["✦ Worker LLM
Evaluator-Optimizer"]
    tools[ToolNode]
    evaluator["✦ Evaluator"]

    person -->|request| stateIn --> worker
    worker -->|tool calls| tools --> worker
    worker -->|text reply| evaluator

    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 worker,evaluator agent
    class stateIn,tools service
    class person actor

Graph edges: START → worker; tools → worker; text path continues on the Evaluator tab.