Territory Grounder · Field guide
Watch the loop, then explore the design
Start with the lab demo if you operate infrastructure. Start with the introduction if you want the idea in everyday language. Each video includes English captions on YouTube.
On a small screen, scroll sideways to see every column.
| Watch | Length | What it establishes |
|---|---|---|
| Three faults on a test machine | 3:57 | Executed repairs and verification in an isolated lab. The featured homepage demonstration. |
| An AI agent turned off the alerts. Three questions to ask | 5:11 | The problem and the three questions: what did it do, was it allowed, did it work? Includes a recorded investigation/proposal and a separate control test, not an executed-repair demonstration. |
| A guided tour, from alert to sealed record | 5:29 | Architecture, modes, gates, verification and the ledger. A design tour with lab screens. |
| DevOps, SRE, AI agents and governance, in everyday words | 7:39 | Background for people new to infrastructure operations. |
| AI agents on the night shift | 5:39 | Opportunities, failure cases and research context. Not a Territory Grounder performance benchmark. |
| Who checks the AI agent? A series, level by level | 1:01 | Trailer for the series. |
Open the complete six-video playlist ↗
Lab transcript
The following is a text walkthrough of the November 1, 2026 lab recording, with links to the relevant chapters. It summarizes the narration; the video’s English captions provide the full transcript.
00:00 — A machine built to be broken
The environment is a disposable test machine. The demonstration asks three questions of each repair: what did the agent do, was it permitted, and did the outcome match? It is a bounded test, not a claim about production reliability.
00:24 — The network link
A network link is disabled. The alert arrives, and the agent examines status, logs and the hypervisor. It proposes restoring the specific link, bound to the machine’s current configuration. A separately computed prediction and the runtime gates precede execution. In this configured lab, the autonomous fallback handles the novelty approval. The link returns and verification records the outcome, about ninety seconds into the recovery.
01:54 — The service
A service is stopped. The agent diagnoses it and starts it through a scoped actuation identity. The repair takes about sixty-two seconds in this run. Repeated stops are flagged for investigation: restarting a process does not explain why it keeps stopping.
02:40 — The virtual machine
The machine is powered off. A deterministic liveness path starts it, with an undo armed until confirmation. This case does not require an LLM. Recovery takes about nineteen seconds in the recording.
03:09 — What did not work first
Earlier attempts exposed a host-label mismatch that needed a fix, a flapping service that the agent refused to keep restarting, and a request requiring an unsupported undo that stayed held. Those cases are part of the demonstration’s limits, not omitted successes.
03:43 — The three answers
The record connects the action, its permission and the observed result. Three selected faults on one test machine demonstrate those paths. They do not supply a fleet-wide success rate. Read the evidence and benchmark limitations.