Every AI automation guru is telling you to build multi-agent systems.
I'm going to tell you what most of them won't: most of you shouldn't.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚠️ The Coordination Tax Nobody's Pricing In
Here's a number that should stop you cold.
In controlled experiments comparing single-agent vs multi-agent setups, moving to multi-agent improved truthfulness by about 28% on Q&A tasks — but at a 3.7× increase in API costs.
Read that again. You pay almost 4× more to close a 3-point accuracy gap.
And Anthropic — the team that actually built one of the most successful multi-agent systems in production — has said the quiet part out loud:
Multi-agent systems use approximately 15× more tokens than standard chat interactions.
15×. Not 1.5×. Not 3×. Fifteen times.
Yet LinkedIn is drowning in "build your 40-agent empire" content. Agency founders are being pitched orchestrated swarms as the next competitive edge.
The math isn't mathing.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
💸 What the Tax Actually Looks Like
Let's break down where the money goes:
🔴 Token blow-up — A three-agent pipeline consumes roughly 29,000 tokens versus 10,000 for an equivalent single-agent approach. If your pipeline doesn't need the specialisation, you're paying 3× for the same result.
🔴 Latency cascade — A four-agent pipeline accumulates roughly 950ms of coordination overhead while actual processing takes 500ms. More time coordinating than working.
🔴 Quadratic coordination cost — Production multi-agent systems don't scale linearly. The effective overhead scales like O(n²) — in tokens consumed, latency added, and cognitive load imposed on the system. Two agents feel fine. Five agents break the bank.
🔴 Compound reliability collapse — Ten sequential steps each at 99% reliability produce only 90.4% overall reliability (0.99^10). Every handoff is a new failure point.
🔴 Debugging hell — In a single-agent system, failures are linear. In a multi-agent system, they're distributed across handoffs. Was it Agent B's logic or Agent C's input that broke the workflow? Good luck.
And it's not just me saying this. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027 — citing escalating costs, unclear business value, and inadequate risk controls as the top three reasons.
That's not a technology problem. That's an architecture problem.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🏗️ What I Almost Built for RFA
Here's the building-in-public part.
A few weeks ago, I was designing the RFA Content Engine — the system that produces this newsletter, the LinkedIn post you probably saw, the Skool community post, the X post, and everything else, daily.
My first design was multi-agent.
📌 Steps 1–3 (topic, research, newsletter draft) → Claude Opus for depth
📌 Steps 4–7 (repurposing, edits, file handoff) → Claude Sonnet for speed
📌 Separate chat sessions, handoff prompts between models, different context per step
On paper? Perfect specialisation. Right tool for each job.
In practice? The coordination tax crushed it.
Here's what actually happened:
❌ Handoff prompts added 15–20 minutes to every session
❌ Context loss between chats meant re-briefing on tone, voice, and earlier decisions
❌ Sonnet couldn't carry the formatting nuance Opus had established — drafts came back bland
❌ I was debugging my own pipeline instead of shipping content
So I killed it.
The new design — documented in v9 of my project instructions — is single-agent. One chat. One model (Opus). One session per day's publishing.
✅ Zero handoff overhead
✅ Full context retained across all 8 workflow steps
✅ Consistency across newsletter + repurposed posts
✅ Faster end-to-end — not because the model is faster, but because there's no friction
The "slower" model in one chat beat the "specialised" setup. Every single time.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🎯 When Multi-Agent Actually Earns Its Keep
Now — am I saying never build multi-agent? No.
Anthropic's research system is multi-agent. It works. They published exactly when it's worth it.
Three scenarios where the math flips:
1️⃣ Genuine parallelisation
Multiple independent subtasks that genuinely run at the same time. Think: research that searches 10 sources simultaneously, or classification across a batch of 500 documents.
The key test: Do the agents need to wait for each other? If yes, you don't have parallelisation — you have sequential coordination with extra steps.
2️⃣ Context that exceeds a single window
Genuinely large research tasks where the input doesn't fit in one prompt. Different agents handle different chunks, lead agent synthesises.
The key test: Have you actually tried compressing the input first? Most "too big for one window" problems are really "I didn't prioritise what to include."
3️⃣ Different models for different job types
GPT-4 for reasoning, Claude for writing, a local model for classification. The coordination here is cheap — just input/output routing. No complex handoffs.
The key test: Are the agents doing genuinely different kinds of work? Or are you splitting the same kind of work across multiple agents just because?
Everything else? Start with one agent and a clear workflow. You can always split later when you hit real limits — not imagined ones.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🧠 The Single-Agent Rule
Here's the test I now apply before building anything multi-agent:
If you can't clearly explain why Agent B can't do Agent A's job, you don't need Agent B.
That's it. That's the rule.
If "Agent B" is just "Agent A with a different prompt" — you don't need two agents. You need one agent with two clearer instructions.
If "Agent B" exists because it felt more sophisticated to have a second agent — you're paying the coordination tax for nothing.
If "Agent B" needs the exact context "Agent A" has — you've created a context transfer problem instead of solving the original problem.
Most teams building multi-agent systems should be building better single-agent workflows instead. Less coordination. More execution. Lower bills.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🛠️ The Playbook
Over the past few weeks, I've been writing down everything I've learned about this — from the RFA Content Engine build, from the research above, from the patterns I've seen agency owners fall into.
It became a playbook:
"The Multi-Agent vs Single-Agent Decision Playbook for Agencies" — 15+ pages covering:
📘 The decision framework (the three questions that tell you which architecture to build)
📘 A copy-paste CLAUDE.md template encoding the single-agent rule in your Claude Projects
📘 Worked examples — including the full RFA Content Engine decision, step by step
📘 The coordination tax calculator (rough numbers for estimating your own multi-agent cost)
📘 When to spawn a sub-agent vs when to stop splitting
📘 Anti-patterns that guarantee a $200/day API bill
If you want it, I'm giving it away free to anyone who asks.
👉 Comment "PLAYBOOK" on today's LinkedIn post and I'll DM it to you.
👉 Skool community members: it's posted directly in the group — join if you're not in yet.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🎯 The Bottom Line
Every agency owner is being told multi-agent is the future.
The data says otherwise. Multi-agent is the future for a specific class of problem. For most agency work — content production, client onboarding, lead qualification, proposal generation — one well-designed agent beats five coordinated ones every single time.
The coordination tax is real. The 3.7× cost premium is real. The 40% cancellation rate is real.
Build the one-agent version first. Earn the right to split.
— Bibhash
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🤝 Let's Talk About It
Have you built a multi-agent system that paid off? Or one that bled money before it shipped?
I want to hear both. Reply to this email or drop it in the Skool community — let's compare notes.
And if you're not already subscribed, the newsletter lands daily with lessons from building RFA in public. Join the agency founders reading it.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📚 Sources
RFA Content Engine Project Instructions v9 (internal, available in RFA Skool community)