Jesse A. Green, Strata Resident Engineer at Palo Alto NetworksPersonal site; not affiliated with or endorsed by Palo Alto Networks.
I build with AI agents, and I check their work.
Every claim, sorted
Counted from the status lists on the project pages, October 2026.
The failures are listed on purpose: checking the agents' work is the method.
Panorama, PAN-OS and Strata Logging Service tooling plus a small AI lab, built with agents. Each project lists what works, what is designed only, and what failed.
If you have two minutes: open the Panorama fleet dashboard, the SLS runbook sample and the hub's Route a request; then read where the checking paid off.
Most numbers here are inputs: tests, reviews, findings. Real-environment proof is still ahead for several of these.
Projects
Every project has a live, read-only demo. Open app starts it in a new tab; nothing you do there reaches my lab or calls a model.
Built for my role
Shown here only with mock or sample data.
-
prototype, pausedWork tool, mock data
Panorama fleet dashboard (panoview)
One read-only view across many Panoramas: a mind map of the fleet, an Attention list, search and exports. Prototype, mock-tested only.
947unit tests, Python 3.9 and 3.134working3designed only1failed or missed
Open appRead the write-upBuilt for a laptop screen. On a phone, see the screenshots in the write-up.
-
built, pausedWork tool, sample data
SLS sizing runbook
A 17-tab Excel runbook, generated from Python, that models Strata Logging Service log volume and per-firewall egress across several Panoramas, with every number tied to a source or an assumption label.
91numbered official sources4working2designed only1failed or missed
-
builtWork tool, placeholders
pan-os-php command builder
A parser that maps the pan-os-php source into a command matrix, an Excel builder with cascading dropdowns, and a cleanup guide that cites the source behind each step.
526actions across 33 config types3working1designed only3failed or missed
The AI lab
Agents, the fence around them, and the machines they run on.
-
running
Agentic AI hub
One tool hub for my AI clients: rules-first routing, per-client MCP tools, and a work queue that agents claim under their own identities.
288parallel claims, exactly one winner per round6working3designed only2failed or missed
-
designed only
AI-zone firewall design
A PAN-OS zone for AI agents on my lab firewall, plus an offline set-syntax checker; designed and staged, not committed.
9,719PAN-OS 11.1 command paths the checker knows4working4designed only1failed or missed
-
running
AI node provisioning
One script turns a consumer laptop into a headless, always-on AI node; two nodes are live.
333 of 333offline tests pass3working2designed only1failed or missed
Also in the lab
Research, documentation and video, built mostly by AI under my rules.
-
dormantDormant proof of concept
DocCompiler
A personal proof of concept: agents wrote user and admin guides for open-source repositories; 516 are online as a read-only library.
2working1designed only3failed or missed
-
pausedPaused; read-only snapshot online
ASTS research desk
A personal research app that keeps sourced facts apart from hype for one satellite stock; a read-only snapshot is online.Read-only snapshot from October 2026. Personal research, not investment advice.
3working1designed only3failed or missed
-
pausedPaused after a Sept-Oct run
RL Recorder
Rocket League replays turned into 4K60 videos with every player's view side by side; 55 public videos.
2working0designed only4failed or missed
At a glance
October 2026| Project | What it is | AI's role / my role | Status | Proof |
|---|---|---|---|---|
| Built for my roleShown here only with mock or sample data. | ||||
| Panorama fleet dashboardOpen appDemo: panoview. | One read-only view across many Panoramas | AI: plan, code, reviews. Me: constraints, two safety switches; the real-environment run is mine (not yet done) | prototype, pausedPrototype, mock-tested only | 947 unit tests; UI smoke test 39 of 39, 0 console errors |
| SLS sizing runbookOpen appDemo: sls. | Excel runbook for log volume and egress, generated from Python | AI: research over public vendor docs, reviews, generator. Me: goal, scope, constraints | built, pausedv1.0 built, not yet used in the field | 91 numbered official sources; 0 Excel error cells after recalculation |
| pan-os-php command builderOpen appDemo: builder. | Command matrix, Excel builder and cleanup guide | AI: parser, workbook, guide. Me: picked the problem from field work | builtWorking build, pinned to one upstream release | 80 command types; 526 actions; 372 named ranges |
| The AI labAgents, the fence around them, and the machines they run on. | ||||
| Agentic AI hubOpen appDemo: hub. | n8n router, MCP tools, agent work queue | AI: code, tests, reviews. Me: routing rules, approval model, veto, live approvals | runningRunning; agents tested on test tickets | Race test: 288 claims, one winner per round; a wrong token gets 403 |
| AI-zone firewall designOpen appDemo: firewall. | PA-440 zone that fences AI agents | AI: design, config, checker. Me: topology, every commit | designed onlyDesign only, not committed | Set-syntax checker (9,719 PAN-OS paths); 350 staged lines, 0 failures |
| AI node provisioningOpen appDemo: nodes. | One script turns a laptop into a headless AI node | AI: research, script, tests. Me: hardware, installs | runningTwo nodes live | 333 of 333 offline tests |
| Also in the labResearch, documentation and video, built mostly by AI under my rules. | ||||
| DocCompilerOpen appDemo: guides. | Agents write user and admin guides for open-source repos | AI: all writing. Me: idea, deploys, prompt-change decisions | dormantDormant proof of concept | 572 repos processed, 516 in the public library; every Admin Guide has a Security section |
| ASTS research deskOpen appDemo: asts. | Evidence timeline and research app for one stock | AI: code, daily reading, reviews. Me: source rules, thesis, every routine prompt | pausedPaused; read-only snapshot online | 293 events, every one sourced; 40 of 120 estimates hit |
| RL RecorderOpen appDemo: rl. | Game replays to every-player 4K video | AI: code, recaps. Me: rules, review of new formats | pausedPaused after a Sept-Oct run | 55 public videos |
Proof is the strongest evidence to date.
About this site
I build tools for my role and run a small AI lab on my own time. Agents write most of the code; I write the spec, set the constraints and verify. Since April 2026 that has run mostly through Claude Code. Most of the volume, such as user and admin guides for 572 open-source repositories (516 of them published in the guide library, the other 56 held back for now), came from scheduled agent runs that worked unattended.
Each page says what I built, how AI was used, and where it stands: what works, what is designed only, and what failed. Most numbers here are inputs: tests, reviews, findings. Real-environment proof is still ahead for several of these.
Work in progress. My lab is partway through a move to a sandboxed, hardened layout, and several workflows are unfinished. The figures show my own lab diagrams, public open-source repos, public game videos, and the day-job tools on mock or sample data only.
How I work with AI
The loop I use for changes now. The hub and the node script went through all of it; earlier projects used parts.
- SpecI write the spec and its "done when" test.
- ResearchAgents research and quote primary sources.
- Skeptic checkSkeptic agents check claims against a file or a source.
- Lens reviewsIndependent reviewers check one lens at a time.
- Tests and my yesOffline tests, a dry run, and a human yes for anything irreversible.
Where the checking paid off
| Case | Agents | Outcome |
|---|---|---|
| Choosing the hub platform | 6 researchers, 8 skeptics and 5 local sweeps | n8n was checked before I built on it |
| Truing up my own docs | 10 auditors, then 10 skeptics | 403 findings: 267 confirmed, 132 corrected, 1 refuted, 3 downgraded; 80 missed items added; 364 edits applied |
| Work queue, before go-live | Race test: 12 rounds x 24 parallel claims | Caught a default match setting that let one agent overwrite another's held row; fixed, then one winner every round |
The truing-up result is the one I point to. The skeptics corrected, refuted or downgraded 136 of the 403 first-pass findings, about a third. They also added 80 that the auditors had missed. First-pass agent output needs a second, hostile read. The full table is on the method page.