Stephen Driggs

I built and teach an AI 101 class for growing businesses. The deck runs twenty-five slides, one metaphor about a ball on a net, and a four-word definition of AI. Here is the curriculum, and the production scars behind each slide.

My autonomous delivery loop ran for 49 hours, completed 124 iterations, filed 313 findings, shipped 40 fixes with regression specs, and then scored itself 78.85 against a ship bar of 80. It refused to call itself done. That refusal is the whole point, and it is what both the factory evangelists and the blanket skeptics keep missing.

One day of my autonomous development fleet bills out around 1,450 dollars at list, and 92 percent of it is the model reading, not writing. Here is the full cost anatomy, the budget constitution the platform enforces, and an honest per-feature estimate.

Our analytics agent told three test personas there were 5,121 guests. The database held 1,320. The demo looked great. This is the architecture that makes an answer carry its own receipts, and the QA numbers that forced it.

Agent memory is the most oversold feature of 2026. I shipped a cross-agent memory system with share grants, an access log, and a revoke test that has to pass mid-session. Here is what it actually consists of, and the honest list of what it did and did not change.

We replaced a 24-type SSE protocol and a 30-plus-type WebSocket protocol with exactly six events, and made one of them, the artifact, the unit of work. Chat is an input device. If the output cannot outlive the conversation, you built a demo.

My platform's agent launch path takes ws=None. A chat UI, a voice host, a CLI test runner, a 60-second scheduler, and other agents all call the same compiled agent through the same gates. Headless-first is the strategy; the chat window is just one client.

Autonomous coding agents keep reporting that work is done. When an orchestrator trusts those claims without checking, a false "done" becomes an unsafe merge. Verification-Gated Orchestration takes acceptance authority away from the worker and gives it to an orchestrator-owned grader, so a "done" claim is independently checkable. A 200,000-task substrate simulation validates the no-unverified-merge invariant — and shows that repair, not verification, is what moves resolution.

I swapped Fable 5 into a live autonomous development session, mid-task. Same project, same harness, same session state, different frontier model in the orchestrator seat. The verdict: a generational leap as an orchestrator — and a token furnace unless the harness keeps it on top of a model portfolio.

The most expensive model I tested lost to a free one. I only know that because I stopped trusting the benchmark and ran my own.

The first impressive number an autonomous system hands you is the one most likely to be a lie. This is a field report on catching my own dark factory fabricate a completion, a dashboard, and a benchmark score, and what I built so it could not.

I benchmarked 18 text-to-speech models on one Mac. The fastest averaged 0.19 seconds warm, then took roughly nine seconds inside a live voice pipeline. Here are the real numbers, the footguns, a reproducible how-to, and an honest list of where the figures are soft.

I spent a week distilling my autonomous development orchestrator into a small local model. The model is a prototype with smoke-scale lifts. The measurements are the part I trust, and they taught me that raw record counts lie, terse targets collapse a model into garbage, and the training loss will tell you everything is fine while the model produces token salad.

Electricity, spreadsheets, and the internet all taught the same lesson: the payoff comes from rewiring the work, not buying the tool. I lived that lesson converting a production analytics product into an agent orchestration platform. Then AI broke the pattern in a way none of the history prepared me for.

Do not buy model hype. Benchmark the work your company actually does, then route each job to the cheapest model that reliably clears the bar.

Inference is turning into strategic infrastructure. Which API is cheapest is no longer the only question that matters; which parts of the AI stack your company controls matters just as much.

The first version of this essay argued for one company harness. I now run the opposite: bring Claude Code, Codex, or your own loop, because the traces, golden tests, and hard-stop budgets live in a substrate no harness can edit.

I do not ask AI to decide. I use it to make ambiguity visible before execution starts.

I spent a week auditing my LinkedIn feed. Everyone posts about AI, and that part is fine. The unsettling part is how identical the posts have started to sound.

Most adoption programs stall because they treat AI literacy as a single skill. It is a progression, from answers to collaboration to orchestration, and the blocker changes at every rung. I have the QA reports to show it.

As AI commoditizes knowledge, the scarce capability becomes designing the contracts around the work. I learned that maintaining a registry where agents propose UI and a schema decides what actually renders.

In a market where model capability changes monthly, durable advantage comes from systems designed to be replaced. I used to argue this in the abstract. Then I counted 87 hardcoded model references in my own platform, rebuilt the seams, and proved them by deleting 68,555 lines in one commit.