Multi-Agent Systems

1. Install Ollama → ollama pull llama3.1:8b (~5 GB — works for agent roles that don't need coding).
2. pip install crewai (CrewAI — role-based multi-agent orchestration).
3. Define a multi-agent team:

from crewai import Agent, Task, Crew, Process
researcher = Agent(role="Market Researcher", goal="Find data on topic", llm="ollama/llama3.1:8b")
analyst = Agent(role="Data Analyst", goal="Analyze findings", llm="ollama/llama3.1:8b")
writer = Agent(role="Report Writer", goal="Write final report", llm="ollama/llama3.1:8b")

research_task = Task(description="Research the top 5 local AI hardware trends in 2026", agent=researcher)
analysis_task = Task(description="Analyze the research findings and identify key patterns", agent=analyst)
writing_task = Task(description="Write a 500-word executive summary with 3 key recommendations", agent=writer)

crew = Crew(agents=[researcher, analyst, writer], tasks=[research_task, analysis_task, writing_task], process=Process.sequential)
result = crew.kickoff()

1. Each agent runs its task, passes output to the next agent. A 3-agent pipeline completes in 1-3 minutes on 12 GB GPU.
2. For debate/consensus patterns: CrewAI supports hierarchical (manager delegates) and sequential (pipeline) processes.
3. Use cases: content creation pipelines, research synthesis, code review teams (reviewer + security auditor + style checker).

Used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb). Runs CrewAI with 3-5 Llama 3.1 8B agents running sequentially. A typical 3-step pipeline completes in 1-3 minutes. For small teams automating content creation, research synthesis, or document review: $400 handles 20-50 pipeline runs/day. Pair with Ryzen 5 5600 + 16 GB DDR4 + 512 GB NVMe. Total: ~$360-405. Multi-agent systems are GPU-efficient — agents run sequentially (one GPU, one at a time) or in parallel on CPU for text-only roles.

Used RTX 3090 24 GB (~$700-900, see /hardware/rtx-3090). Runs CrewAI with 5-10 agents using Llama 3.3 70B Q4 for reasoning-heavy roles (analyst, strategist) and 8B models for simpler roles (researcher, formatter). For enterprise multi-agent systems (legal document review with 3 specialized agents, financial analysis with analyst + auditor + writer): the 70B model dramatically improves reasoning quality in complex agent roles. Total: ~$1,800-2,200. For parallel multi-agent execution: 2× RTX 3090 runs two 70B agents simultaneously.

The mistake: Creating a 10-agent system where every agent is the same model with a different system prompt, expecting dramatically better output than a single agent. Why it fails: Same model + same knowledge cutoff + different prompt = the same agent in cosplay. The "researcher" and "analyst" both have identical knowledge and reasoning patterns — the "analyst" can't find insights the "researcher" couldn't. The output is just the same answer rewritten 3 times. The fix: Use different models for different roles. Researcher = 8B (fast, broad search). Analyst = 32B (deeper reasoning). Writer = 8B (fluent output). Different model architectures bring different strengths. Or: give agents access to different tools/data — the researcher has web search, the analyst has a database, the writer has style guides. Differentiated inputs + differentiated models = true multi-agent value. Same model + same data = one agent with extra steps.