Home / Series
Watch AI agents build a live show on Google Cloud
An agent builds and runs a panel show performed by AI models, on its own Google Cloud project, while the operator plans, directs, reviews, and holds the keys. It starts on paper, goes up in slices, and was filmed as it happened, problems included.
What this series built
The Fifth Circle
Five AI models argue a viewer's question while Hal rules. The show runs on its own channel, from its own Google Cloud project, and publishes its own episodes. Every part of it was built in the series below.
Episodes
-
01The plan before the VM: 27 hours, 15 artifacts, not one machine turned on
Before any machine was turned on, the build was a folder of planning documents: about twenty-seven hours of work and fifteen artifacts. This walks through each one and what it settled, from the build sheet and the decisions register to the cast, the mock episode, and the audition harness, and explains why the build starts without a rehearsal.
-
02Giving an AI agent a Google Cloud project without handing it the keys
The first evening of the build, on camera. An agent creates a cloud project in its own folder and on its own billing, takes an identity it borrows through service-account impersonation instead of holding a key, and sets a budget that detaches billing at the cap, proven with a synthetic alert first. It also covers the setup lessons that cost the most time: which shell the agent works in, and why credentials live in a context.
-
03An AI agent sets up DNSSEC on Google Cloud DNS, start to finish
An agent moves two new domains onto Google Cloud DNS with DNSSEC on from the start, then stops where its reach ends: the registrar account is personal, so a second agent sets the nameservers and DS records there with a human approving each save. Every step is then proven from outside the system, from the delegation and signatures to a full DNSViz chain of trust, and the zone gets its first real record. It also covers why the registrar's one-click DNSSEC toggle is never used.
-
04Protecting an AI agent's inbox from spam floods and prompt injection
The show takes letters from the public, and every letter a model reads costs money. This builds that front door and then attacks it: the public's input is walled off, a scanner reads every letter before any model does and is tested against a deck of hostile letters, a human check guards the form, and 1,500 letters are fired at the mailbox until storm mode trips and nothing reaches a model. The design is shown; the detection rules and thresholds are not.
-
05When AI agents do the work, what's your job?
Over five working sessions an agent stood up a complete system on cloud infrastructure, and the operator never opened the virtual machine. This talk piece covers what the operator's job became instead: holding the credential boundary, steering at the load-bearing forks, challenging "done," inspecting what you expect, delegating overnight, and keeping the record that lets the work stop and resume.
-
06Securing an AI-agent system on Google Cloud, layer by layer
The security of an agent-built cloud system, in the order it was built: the box and its blast radius, one key per purpose, budgets that actually stop spending, and a network with no way in and one way out. It also covers why permission errors in testing mean the design is working, what a public surface really exposes, and why recovery is part of security.
-
07AI agents built a YouTube show that publishes its own episodes. Here's the system
Six working sessions after the plan, the system argues each question live and publishes the show to its own channel. This wrap-up covers the final security sitting, the component that was never built, a kill switch tested by actually revoking the keys, what the first episode cost, and review before the upload.
-
08My AI agents said the build was done. Here's what was missing.
The build was declared finished, and the first real request sent through it found that the website's front door had never been wired. This is the day after done: a frontier model audits what the helper agents dropped, the trace turns up six gaps, one question uncovers a missing feature, a pick goes wrong and gets rolled back, and the first viewer ticket goes end to end. The next morning the schedule ran with nobody touching it.
-
09A plan is a budget for change: what 27 hours of planning bought an AI-agent build
What twenty-seven hours of planning bought the build, and what to do on the day the plan breaks anyway. A talk piece on plans as budgets for change: plan in the big model and execute in the middleweight, size every phase to one context window, write the irreversible moments into the plan so the agent stops and the human decides, and the week a helper agent quietly kept a prototype nobody asked for. Plans change; quiet changes don't.
