Agent Workflow¶

This page stops looking at Agent as a component and focuses on how one conversation turn advances internally. In other words, this page is about behavior flow, not object structure.

1. Core workflow idea¶

Dubbo Admin AI currently uses a ReAct-style loop that breaks a complex question into:

Think: understand the problem and decide whether external capabilities are needed
Act: call tools or retrieval
Observe: merge the results and generate intermediate summaries or the final answer

This design is better than a single monolithic prompt because it gives the system a controlled window for calling external capabilities.

2. Default workflow diagram¶

flowchart TD Start["Receive user input"] --> Think["think"] Think --> Act["act"] Act --> Observe["observe"] Observe --> Decision{"Has FinalAnswer?"} Decision -- No --> Think Decision -- Yes --> End["Output final answer"]

3. What each stage does¶

think¶

Understand the user question
Determine the problem type
Choose potentially useful tools
Produce the next-step action suggestion

act¶

Decide whether to actually call tools based on the think result
Execute tools and collect outputs
Produce structured ToolOutputs

observe¶

Read tool results
Generate an intermediate summary
Stream content to the user progressively
Produce the final answer when conditions are met

4. Termination conditions¶

The loop does not run forever. The current main stop conditions are:

observe produces FinalAnswer
max_iterations is reached

max_iterations is an important safety guard to avoid infinite loops where the model keeps trying to call tools without obtaining useful results.

5. Where streaming output is produced¶

Real streaming output is produced in the streaming flow inside the observe stage. It reads stream chunks from the model and pushes text to Server through UserRespChan.

sequenceDiagram participant Model participant Observe participant Channel participant Server participant Client Model-->>Observe: stream chunk Observe-->>Channel: StreamFeedback(text) Channel-->>Server: UserRespChan Server-->>Client: content_block_delta

6. How context enters the workflow¶

The Agent does not feed every stage only the raw user input. It also injects history-window messages into the prompt. That context comes from the Memory component and is read by sessionID.

So workflow quality depends not only on prompts and models, but also on:

whether the session is reused
whether the history window is reasonable
whether turns advance correctly

7. Why tools are not called every time¶

The Agent workflow does not treat tool invocation as the default path. In the current implementation, act only tries to call a tool when think decides the problem is not ordinary Q&A and suggests a tool to use.

That avoids two issues:

simple questions taking heavy external calls and wasting time and cost
tools being overused by the model and making the loop unnecessarily complex

8. The most effective way to observe this workflow¶

During production troubleshooting, inspect logs and latency by stage:

whether think produced a structured judgment successfully
whether act actually issued a tool request
whether observe started streaming output
whether the workflow ends steadily after a certain iteration

If you only read the final answer, it is hard to tell whether the problem is "the model cannot reason", "the tool is not called", or "the result is not integrated correctly".