[MAICE Dev Log 2] How the agents collaborate (Multi-Agent System)

1. Why one LLM was not enough

At first, we tried putting everything into one huge prompt: pedagogy constraints, curriculum terms, student-level adaptation, and response policies.

The result was predictable: instructions were missed or conflicted.

So we moved to separation of concerns. Question analysis and answer generation had to be split into different specialized agents.

2. Reliable communication with Redis Streams

In MSA environments, communication reliability is critical. We chose Redis Streams over plain HTTP so messages are not lost even if an agent restarts.

async def read_from_backend_stream(self, count: int = 1, block: int = 1000):
    messages = await self.redis_client.xreadgroup(
        self.agent_consumer_group,
        self._consumer_name,
        {self.BACKEND_TO_AGENT_STREAM: ">"},
        count=count,
        block=block,
    )

This design let the system resume safely after container restarts.

3. Question-processing pipeline

sequenceDiagram
    participant S as Student
    participant BE as FastAPI Backend
    participant QC as QuestionClassifier
    participant AG as AnswerGenerator
    participant QI as QuestionImprover

    S->>BE: "I don't get this"
    BE->>QC: Redis Stream (Classify)
    QC->>QC: Analyze intent and Bloom level

    alt Needs Clarification
        QC-->>BE: Needs_Clarification
        BE->>QI: Improve task
        QI-->>BE: Clarification question
        BE-->>S: Send clarification question
    else Answerable
        QC-->>BE: Answerable (e.g., K3)
        BE->>AG: Generate with K-context
        AG-->>BE: Structured answer
        BE-->>S: Deliver answer
    end

This is where research logic maps into runtime behavior.

4. Failures we had to solve

Deadlock risk: solved with hop count and cycle detection.
Context discontinuity: solved with shared session context across agents.

5. Why specialization mattered educationally

The main gain was not just cleaner architecture. It was more consistent pedagogical behavior:

clearer routing between K1-K4 response structures
better handling of unclear questions via clarification loops
reusable learning summaries through the observer layer

Detailed statistical evidence is covered in Post 7.