I’m building a fully autonomous B2B support pipeline. I’m choosing a single-prompt huge-context LLM script over a complex Multi-Agent

[Foltern]

Hey everyone, I’m finalizing the architecture for an automated customer support and triage system for a mid-sized SaaS company.
The Constraints:

• Volume: ~5,000 incoming tickets/emails per day.
• Goal: Autonomously resolve or perfectly draft replies for at least 60% of tickets, including processing refunds, explaining docs, and escalating bugs.
• Accuracy: We absolutely cannot have the AI hallucinate a refund policy or promise a feature we don't have.

The Proposed Architecture (What I'm choosing):
I'm ditching agentic frameworks entirely. I’m using a simple, sequential Python script with a Tier-1 LLM (Claude 3.5/GPT-4o/Gemini Pro) utilizing its massive context window.

• When a ticket comes in, a basic Python script pulls the user's entire CRM history, billing data, and the last 10 interactions via API.
• We dump all of this raw JSON data, plus our entire 50-page markdown internal knowledge base, directly into a single, massive system prompt using Context Caching.
• The LLM is forced to output a strict JSON response containing the action_to_take and reply_draft.

The Rejected Architecture (What I'm avoiding):
I explicitly rejected the popular "Swarm" or Multi-Agent frameworks (LangGraph, AutoGen, CrewAI, etc.).

• Why I hate it for this: Building a "Classifier Agent" that talks to a "Retrieval Agent" that hands off to a "Drafting Agent" feels like a fragile, over-engineered mess. It multiplies latency, compounds hallucination risks at every handoff, and is an absolute nightmare to debug when the agents get stuck in a reasoning loop. Why build 5 agents when one smart model with a 1M+ token context window can just read the whole situation at once?

The Ask: Destroy My Design
I know dumping 200k tokens of raw data into a prompt is the "lazy" way out, but with context caching, it feels far more robust than a brittle multi-agent state machine.
As automation experts, tear this apart.

1. Where is my blind spot?
2. When this system breaks in production, what specifically causes it? (Attention degradation? Context overload?)
3. Prove to me why a multi-agent framework or a complex RAG pipeline is actually necessary here.

feedbacks are much appreciated