Table of Contents
The AI industry has it backwards. They’re selling agents as the product. Agents aren’t the product. Agents are the tool you use to build the product.
The product is workflows.
I’ve been building AI agents for two years. Marketing agents, financial agents, legal agents, developer agents, recruiting agents — over a dozen specialized agents across multiple platforms.
I believed in this deeply enough to publish “The Agentic Startup Manifesto” in June 2025.
Today I’d write a different manifesto.
The Agent Fantasy
The industry loves agents because they map to a comforting metaphor: digital employees. A robot that does your job. The pitch writes itself.
I quoted that pitch. Two years and too many wasted tokens later, here’s what I actually found.
What Breaks
Conversational interfaces — having an LLM decide what to do next — carry every problem of human delegation, minus the things that make humans reliable.
It’s hard enough to communicate a product requirement to an engineer who remembers your last conversation and asks clarifying questions. Now imagine that the engineer forgets what you said five minutes ago. Never asks questions. Just assumes.
That’s an AI agent. That’s what hallucinations actually are — confident assumptions with no memory.
We built a multi-agent platform where any employee could ask anything about the company. The result: no one knew what to ask. And when they did, they didn’t know what to do with the answer. A shiny new toy.
Every agent call requires LLM inference. Workflow paths change between runs. You can’t reproduce results. You can’t observe what happened. You can’t run them in parallel. Each conversation is a new roll of the dice — on your budget.
What Works
Then we built something simple. An LLM-powered workflow to analyze meeting transcripts and generate structured notes. No agents running in production. No conversations. A deterministic pipeline with intelligence injected at the right step.
Someone who spent an hour watching a recording now spends five minutes reading the notes. Not the generic AI-generated summaries you’ve seen (Zoom?) — a fine-tuned workflow that captures decisions, action items, and context the way your team actually talks.
Now extend that principle. Apply the same pattern — automation + intelligence — across your entire organization. No one has to ask. No one has to coordinate with an agent. The LLM does the thinking. The workflow does the moving.
Workflow-First AI
Here’s the framework I didn’t have two years ago:
Agents are builders. Workflows are the product.
I use AI agents every day — Cursor, Copilot, LLM-powered development tools. They’re fast. A developer who understands prompting, context, and the right toolchain can ship in days what used to take months. I’ve done it. Repeatedly.
But here’s what the industry gets wrong: they think agents ARE the product. They deploy agents to run business operations — conversational interfaces orchestrating themselves, deciding what to do next, burning tokens on coordination.
That’s the wrong job for an agent.
The right job for an agent is to build the automation. Write the code. Generate the pipeline. Create the workflow. Then get out of the way.
The workflow runs in production. Deterministic. Observable. Reproducible. Cheap. No LLM deciding what to do next. No results that change between runs. No coordination overhead.
The question isn’t “what agents do we need running our business.” The question is “what workflows can we build — using agents — to automate our business.”
The first question leads to science projects. Chatbots nobody uses. Multi-agent swarms that eat your AI budget on coordination.
The second question leads to compounding automation. Every workflow you ship runs forever. Every decision point where you inject intelligence gets better with data. No coordination tax. No lost context. No hallucination chains.
Connect enough of these workflows, and you don’t have a collection of tools. You have an AI system running your company.
Agents are builders. Workflows are the product.
- AI Moved the Bottleneck to Your Head - 02/19/26
- Stop Building AI Agents - 02/17/26
- From SaaS to AI Agents - 05/27/25