[Dicsussion] Loop Enginnering In AgentScope

[DavdGao]

Background

We have been integrating various loop engineering capabilities into AgentScope. However, the current implementations are relatively scattered, including:

• A budget-aware middleware that breaks the loop when the token budget is exceeded
• [PR] A mechanism that provides hints to the agent and performs a fallback when tool calls fail consecutively
• [PR] Support for ReAct-style loops

This raises the question of whether we should provide a unified agent middleware — e.g., a LoopMiddleware — to consolidate these loop-related functionalities. However, there is an inherent trade-off: what should live inside the agent class itself, versus what should be implemented as middleware?

To address this, we are opening this discussion. The loop engineering capabilities currently under consideration are as follows:

---

Budget Control

• Token budget (total or per-reply)
• Time budget (total or per-reply)

Goal

• Achieve a specific objective, such as successfully running a unit test or executing a target function

Self-Awareness

• The agent is aware of its current state, including elapsed time, context length, etc.

---

On Implementation Placement

Personally, I think capabilities that universally enhance agent performance should be implemented inside the agent class, while capabilities that are function-specific should be placed in the middleware layer, allowing developers or users to configure and enable them selectively.

On Implementation Approach

One key concern is whether we should establish a unified implementation channel. Given that different middlewares may provide loop-related functionality from different angles, having each middleware inject a HintBlock into the agent context is arguably not an ideal approach.

---

We welcome any suggestions regarding functionality, implementation placement, and implementation approach — please feel free to share your thoughts in this discussion!

[Mrart]

I think we also need to add these two features：

1. We should also implement stagnation detection (or progress monitoring). This mechanism will identify when an Agent thread becomes unresponsive or stalls for an extended period without producing the next result, and automatically terminate the Agent.
2. Let's tailor our error messages for the Agent, not for humans. We should structure them so they are easy for the Agent to parse and understand, instead of just being friendly for people to read.

[Mrart]

Additionally, in loop scenarios, tools may be invoked multiple times. Therefore, ensuring idempotency across these repeated calls is also crucial.