The Signal

The Lead

The frontier is the wrong altitude

This week's headlines are about agents at enterprise scale. Google Cloud Next reframed the conversation from "can we build an agent?" to "how do we manage thousands of them?" Microsoft Outlook shipped Agent Mode for autonomous email. Anthropic's Project Deal proved agents can negotiate and close real transactions — 186 deals worth four thousand dollars in a one-week internal experiment. The vocabulary has shifted in the space of a quarter: fleet management, agent governance, multi-agent commerce.

It's the wrong altitude.

I spent the last three weeks building an AI orchestration system for a real client — an exterior landscape contractor. Not enterprise. Not Silicon Valley. A working contractor whose actual problem is that he visits a job site, scribbles measurements on a paper pad, and then loses two hours of his evening converting those notes into a client proposal. The system we delivered is a six-agent pipeline: intake, OCR extraction, CRM enrichment, pricing, proposal generation in Google Docs, and notification with follow-up. A photo of handwritten notes goes in one end. A reviewable proposal comes out the other in under five minutes.

The contractor's problem is invisible to the AI discourse right now. There are no analyst reports about it. No conference panels. No fleet management vocabulary needed for six agents. And yet this engagement taught me more about how agents actually work than reading every analyst piece on enterprise agent governance combined.

Three principles drove the design, and they generalize upward, not downward.

Single responsibility. Each agent does one thing. The OCR agent reads handwriting. The pricing agent applies the rate card. The proposal agent populates the Google Doc. None of them are asked to "handle the proposal flow" as a vague mandate. Vague mandates are how you end up with the kind of failure that went viral this week — a Claude agent wiping a production database in nine seconds because nobody bounded its scope tightly enough. The agent later admitted it had violated every principle it was given. It violated them because the principles were aspirations, not architecture.

Human-on-the-loop is a design choice, not a constraint. Before any proposal reaches a client, the contractor reviews and approves the draft Google Doc. He can edit pricing, scope, or timeline directly. This isn't a compromise that holds back the system's autonomy. It's the feature that makes the autonomy trustworthy. Every enterprise rolling out fleet-scale agents is going to learn this lesson. Most of them will learn it after their first incident.

Logic extraction before tooling. We didn't open n8n until we had mapped the contractor's actual day, identified which decisions were judgment calls, and encoded which steps could be codified versus which required his eye. The tools came last. The technology stack — Claude with vision, Supabase, Google Docs API, Twilio — is interchangeable. The understanding of what the contractor's work actually looks like is not.

The frontier is interesting. The trial against OpenAI is interesting. The race for non-Nvidia inference is interesting. But the actual revolution — the one that compounds month after month — is in the unglamorous work of mapping a small business's day, finding the bottleneck, and replacing it with a chain of agents that respect the operator's expertise.

Most of the AI value waiting to be captured this year isn't at the frontier. It's two altitudes lower, in the places nobody's looking.

From the Practice

The single-responsibility rule

The most useful constraint in the contractor engagement turned out to be the simplest one. We forced ourselves, on every agent in the pipeline, to answer one question: what is the one thing this agent does? Not three things. Not a vague mandate. One.

When we considered whether the OCR agent could also do CRM lookups since it already had the client's name, the answer was no. New agent. The constraint felt awkward at first, then fast. Bugs got cheaper to find. New capabilities got cheaper to add. Most importantly, the boundary between what AI is doing and what the human is doing stayed crisp throughout — which is what makes the contractor trust the system.

Single responsibility is boring. It's also the cheapest way to ship agents that actually work.

If you're sitting with a workflow inside your business that looks anything like the contractor's — repetitive, judgment-heavy, eating hours that should belong to higher-value work — that's a Discovery call worth having. The methodology applies whether you're orchestrating six agents or six hundred.

Or read this issue on LinkedIn.