Agents

The Agent’s role is to decide which model to invoke. It can be more independent than a simple chat or reasoning model, and can orchestrate multiple models and tools in a single workflow.

Orchestration flow

This is our current vision for the Agent workflow. It is not yet implemented, but it will be the basis for the final design.

graph TB
    subgraph User
        U["Web Interface"]
    end
    U -->|prompt| A["Particula Agent"]
    A -->|vector search| V[(Vector Store)]
    V -->|context| A
    A -->|choose model| M{"Chat/Reasoning"}
    M -->|LLM response| A

    %% Python call delegated to a Chat model
    A -->|invoke| CP["Simulation Chat Model"]
    CP -->|generate simulation| PT[[Python Tool]]
    PT -->|run simulation| SC[[Simulation Calculation]]
    SC -->|results| CP
    CP -->|return analysis| A

    A -->|final answer| U

Decision policy:

• quick explanations, short multimodal queries → Chat (GPT‑4o‑mini)
• medium complexity with images or ≤128 k context → Chat (GPT‑4o / GPT‑4.1mini)
• long context, code generation, or tool execution → Reasoning (o3mini / o4‑mini)
• deeply nested logic or heavy planning → Reasoning (o3 or larger o‑series)

Note: This Agent workflow is our goal.

---

Reference: OpenAI Agents Docs