J.P. Morgan showed how they built their multi-agent research system, and it looks almost exactly like the framework I use across my businesses. Here’s the rule, the 4-part architecture, and how to run it as a solo operator.
Most beginners build one giant agent and try to make it do everything. It gets confused, forgets things and makes mistakes constantly.
The rule is one agent, one job, like an employee with one specialty. Build a stack of those, then build one supervisor that ties them together. J.P. Morgan calls that a supervisor agent. I call mine my co-founder.
Here’s the actual setup for my e-commerce business. Each agent only knows what it needs for its one job.
01
Email marketing agent
Writes the weekly emails, manages sequences and tracks opens and clicks. (How I set it up)
03
Meta Ads agent
Same job as #2, for Facebook and Instagram. Different platform, different agent. (The guide)
04
Customer service agent
Sorts incoming questions, drafts replies and flags refunds for a human to approve.
05
Daily reporting agent
Pulls numbers from all of the above into one morning report.
06
The co-founder (supervisor)
Watches the others, tells me when one breaks and goes in to fix it. (Build your own)
J.P. Morgan’s Private Bank built a research assistant called Ask David (Data, Analytics, Visualization, Insights and Decision-making) and walked through it at LangChain’s Interrupt conference in May 2025 (watch the talk). It has a supervisor and three specialist agents, and the same shape works for anything you’d build.
01
Supervisor agent
The one you talk to. It figures out what you want, sends the work to the right specialist, keeps short and long-term memory, and knows when to pull in a human.
02
Structured data agent
Turns plain-English questions into SQL queries or API calls, runs them and summarizes the answer. Use this for anything that lives in a database or a SaaS tool like Shopify or your ad accounts.
03
Unstructured data agent (RAG)
Searches documents like emails, meeting notes and presentations that have been indexed for search, finds the right passage and returns it. This is what makes “answer using my company’s docs” work.
04
Analytics agent
Does the math. Simple questions go straight to an API. For complex ones, it writes code, and J.P. Morgan has a human supervise running it. The supervisor calls this agent for calculations instead of lookups.
You don’t need LangGraph or a team of engineers. Here’s the same pattern in a small-business stack.
01
Build each agent with Claude Code
Describe what the agent does and Claude Code writes it. Give every agent its own folder, with its own instructions file and its own credentials.
02
Run them in OpenClaw
OpenClaw is a free, open-source agent you host yourself on a computer or server that stays on. It runs jobs on a schedule and can keep separate agents with their own settings. To use Claude inside it, you can reuse your Claude Code login on the same machine, which counts against your plan’s usage limits, or add a pay-per-use Claude API key. OpenClaw recommends the API key for automation that runs all day. (My OpenClaw setup guide)
03
Talk to them in Telegram
Give each agent its own Telegram bot, or have one supervisor bot route to all of them. My morning report, refund flags and ad alerts all land on my phone. No dashboard.
01
Pick the most painful job
The one that eats your time every week. That’s agent #1. For most people it’s customer service or daily reporting.
02
Spec it, then build only that one
Paste this into Claude first. It interviews you and writes a one-page spec covering the trigger, inputs, output, permissions and what needs your approval. Then hand the spec to Claude Code. Never add a second job to that agent.
03
Don’t add #2 until #1 works
Watch it run for a week and fix what breaks. Then build the next one. The temptation is always to scale before it’s stable. Resist it.
Honest Limitations
Without a supervisor checking the others, errors pile up quietly. The supervisor needs guardrails too, so it never spends money or posts anything public without your approval. And API costs add up. Five agents running daily is manageable, but a supervisor that checks everyone every few minutes will burn tokens fast, so check on a schedule instead. For the full rules I give my agents, grab my guardrails template.