← sources

Multi-agent system burns 4-5 LLM calls per task and I keep hitting Groq's free-tier limit. How do people handle this?

source: reddit (reddit:AI_Agents)published 2026-10-02 16:23ingested 2026-10-03 02:35

retained text

I'm building a retail automation system — 1 orchestrator + 4 specialist agents (inventory, finance, email marketing, customer care) — running openai/gpt-oss-120b on Groq's free tier.

The problem: a single user action burns 4-5 LLM calls.

Example — user says "10 units of jeans sold":

1. Orchestrator parses the intent

2. Handoff to inventory agent → deducts quantity from DB

3. Handoff to finance agent → generates the invoice

4. Back to orchestrator

5. Orchestrator formats the final response

Five calls for one sentence. A few of these in a row and I hit the rate limit, and the whole thing stalls until the window resets.

What I'm trying to figure out:

\- Do people actually reduce the call count, or is the honest answer "you pay once you're past prototype"?

\- Steps 2 and 3 feel deterministic — a DB update and an invoice template don't obviously need an LLM. Is the right move to pull those out of the agent entirely and let the orchestrator call them as plain functions?

\- Does anyone merge multiple agent steps into one call? Does that hurt reliability?

Happy to share the orchestration setup if it helps.

document fetched 2026-10-03 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wvy6b7/multiagent_system_burns_45_llm_calls_per_task_and