SYSTEMS
How to Structure Your AI Agents 📚
JPMorgan published their multi-agent system. It looks almost exactly like the framework I use across my businesses. Here’s the rule, the 4-part architecture.
The One Rule
Most beginners build one giant agent and try to make it do everything. It gets confused. It forgets things. It makes mistakes constantly.
The rule: one agent, one job. Like an actual employee with one specialty. Build a stack of those, then build one supervisor that ties them together.
JPMorgan calls this exact pattern a supervisor agent. I call mine my co-founder. Same architecture.
5 Agents + 1 Co-Founder
For my e-commerce business, here’s the actual setup. Each agent only knows what it needs to do its one job.
Agent 01
Email Marketing Agent. Writes weekly emails, manages sequences, tracks opens/clicks.
Agent 02
Google Ads Agent. Bid management, keyword adds, ad copy variants, daily budget alerts.
Agent 03
Meta Ads Agent. Same job as #2 but for Facebook/Instagram. Different platform, different agent.
Agent 04
Customer Service Agent. Triages incoming questions, drafts replies, flags refunds for human review.
Agent 05
Daily Reporting Agent. Pulls data from all of the above into a single morning dashboard.
The Supervisor
Co-Founder Agent. Watches the others. Tells me when one breaks. Goes in and fixes it.
The 4-Part Architecture
JPMorgan’s multi-agent system — nicknamed Ask D.A.V.I.D. — was presented at LangChain Interrupt 2025. It has four core agent roles. They map directly to anything you’d build at home.
┌────────────────────┐ │ SUPERVISOR AGENT │ ←── you talk to this one │ (orchestrator) │ └─────────┬──────────┘ │ ┌─────────────┼─────────────┐ ↓ ↓ ↓ ┌──────────┐ ┌──────────┐ ┌──────────┐ │STRUCTURED│ │UNSTRUCTURED │ANALYTICS │ │ DATA │ │ DATA ││ AGENT │ │ │ │ (RAG) ││ │ │ SQL/APIs │ │ emails, ││ runs │ │ │ │ notes, ││ models / │ │ │ │ PDFs ││ code │ └──────────┘ └──────────┘ └──────────┘
The one you talk to. Understands intent, decides which sub-agent to call, holds short and long-term memory, escalates to a human when needed.
Translates natural language into SQL queries or API calls. Runs them. Summarizes the result. Use this for anything in a database or a SaaS API.
Vectorizes your emails, meeting notes, PDFs, audio transcripts. Finds the right snippet. Returns it. This is the agent that makes “answer using my company’s docs” possible.
Runs the actual computations — financial models, simulations, custom code. The supervisor calls it for anything that needs math, not just retrieval.
How To Apply It With No Engineering Team
You don’t need LangGraph or a 50-engineer team. Here’s the same pattern in a small-business stack.
Way cheaper than enterprise agent platforms. You describe what each agent does, Claude builds it. One folder per agent.
Hosts your agents and runs them on schedules. Each agent gets its own runtime, its own credentials, and its own logs. Cheaper than enterprise platforms; sturdier than running scripts on your laptop.
One Telegram bot per agent (or one supervisor bot routing to all of them). I get my morning report, refund flags, ad alerts — all on my phone, no dashboard required.
Don’t Build All 5 Today
The one that bleeds your time every week. That’s your first agent. For most people it’s customer service or daily reporting.
Spec it: input, output, schedule, what tools it can call. Build with Claude Code. Don’t add a second job to it.
Watch it run for a week. Fix what breaks. Then build the next one. The temptation is always to scale before you’ve stabilized. Resist.
In outbound marketing and career positioning, generic applications have near-zero conversion. Treat yourself as a high-ticket solution: identify the company's pressing operational pain points, build a mini-audit or work sample using AI before you apply, and bypass crowded channels by reaching out directly to the decision-maker with structured value.