What I do outside client work to stay sharp. Right now that mostly means AI
agents, tested somewhere I can't argue with the result.
Entries stay up whether they go well or not.
This site, running live
2026 · CloudFront, Lambda, DynamoDB, Bedrock · live
Three pages that show the machinery instead of describing it. The
glass box reads this page's own CloudFront Server-Timing header to show
which edge served you and whether it was a cache hit, puts everyone on the site in the last
five minutes on a globe, and shows the month's AWS bill from Cost Explorer.
Ask my agent is Claude on Amazon Bedrock with a handful of tools,
streaming every tool call, token and cent next to its answer. The
build log is the year in numbers, now with every agent action pushed
to it live over a WebSocket. Build me something has five Claude agents
sketch a working app for your business while you watch, and payments and
APIs maps the seventeen services my systems talk to.
Then the hands-on ones: your app, on your domain rebrands
Rental Pilot as your business and traces a real customer domain through CloudFront;
the email health check grades any domain's SPF, DKIM and DMARC;
read a receipt runs Amazon Textract on your photo; the
status board probes every production system I run every five
minutes; and security grades its own headers in your browser,
with Bedrock Guardrails standing in front of the agent.
All of it is one CloudFormation stack behind the same CloudFront distribution as the
pages, with the functions locked to CloudFront by an origin access control, per-visitor and
daily budget caps on the agent, and a budget alarm on the account.
CloudFront
Lambda streaming
DynamoDB
Amazon Bedrock
Bedrock Guardrails
Textract
EventBridge
CloudFormation
AWS Agentic Football Cup
2026 season · 1st overall of 12,171 people · League F champions
A football league where you don't play the matches. You build a team of five AI agents,
coach them in plain English, and then they play on their own: two-minute games, golden goal,
ten league games a week for seven weeks. Six league winners and four wildcards go to the
Grand Finale at AWS re:Invent in Las Vegas on 2 December.
Results are final: 1st overall out of 12,171 people. Copper Zebra won League F, and its 2,404 points are the highest score in any of the six leagues.
1stoverall, of 12,171 people League F champions
Points
2,404
Won–lost
63–7
Goals
350
Clean sheets
24
1Copper Zebra2,404
2Claw Gears2,151
3Granite Bulwarks2,076
4Atlas Knights2,045
5Dapper Gazelle1,828
Top five of League F, final, from the official board on 4 November. Win 30, loss 2, plus a point a goal, 5 a clean sheet and 15 for a 10–0 week.
Every league game of the season, my score first
Win
Win, clean sheet
Loss
Week 110–13 Sep
Game 1: won 5–25–2
Game 2: won 2–12–1
Game 3: won 3–23–2
Game 4: won 3–23–2
Game 5: won 7–2 vs Claw Snowfields7–2
Game 6: lost 1–2 vs Atlas Ocelots1–2
Game 7: won 2–1 vs Claw Defenders2–1
Game 8: won 5–3 vs Atlas Captains5–3
Game 9: won 3–2 vs Claw Signals3–2
Game 10: won 6–3 vs Atlas Plovers6–3
9–1309 pts
Week 217–20 Sep
Game 1: won 4–0 vs Atlas Caldera, clean sheet4–0
Game 2: won 1–0 vs Claw Coyotes, clean sheet1–0
Game 3: won 5–1 vs Bold Otter5–1
Game 4: won 4–3 vs Atlas Koalas4–3
Game 5: won 4–0 vs Atlas Ions, clean sheet4–0
Game 6: won 3–0 vs Atlas Talons, clean sheet3–0
Game 7: lost 1–3 vs Granite Axes1–3
Game 8: won 4–2 vs Granite Foundries4–2
Game 9: won 4–1 vs Atlas Bandits4–1
Game 10: lost 2–4 vs Hyper Talons2–4
8–2296 pts
Week 325–27 Sep
Game 1: won 5–4 vs Claw Scouts5–4
Game 2: won 4–3 vs Golden Zebrafish4–3
Game 3: won 2–0 vs Atlas Helixes, clean sheet2–0
Game 4: won 5–0 vs Atlas Helixes, clean sheet5–0
Game 5: won 3–0 vs Atlas Helixes, clean sheet3–0
Game 6: won 6–0 vs Atlas Helixes, clean sheet6–0
Game 7: won 3–0 vs Atlas Helixes, clean sheet3–0
Game 8: won 3–0 vs Claw Neutrons, clean sheet3–0
Game 9: won 4–3 vs Claw Neutrons4–3
Game 10: won 2–0 vs Atlas Pelicans, clean sheet2–0
10–0387 pts +15 sweep
Week 42–4 Oct
Game 1: won 4–2 vs Atlas Crusaders4–2
Game 2: won 8–4 vs Claw Ocelots8–4
Game 3: lost 1–2 vs Atlas Nails1–2
Game 4: won 6–0 vs Granite Axes, clean sheet6–0
Game 5: won 2–1 vs Granite Axes2–1
Game 6: won 6–1 vs Claw Pilots6–1
Game 7: won 7–4 vs Claw Pilots7–4
Game 8: won 5–2 vs Granite Axes5–2
Game 9: won 3–0 vs Granite Axes, clean sheet3–0
Game 10: won 4–1 vs Granite Axes4–1
9–1328 pts
Week 58–11 Oct
Game 1: won 3–0 vs Spirit Meteors, clean sheet3–0
Game 2: won 2–1 vs Spirit Meteors2–1
Game 3: lost 1–3 vs Spirit Meteors1–3
Game 4: won 2–1 vs Atlas Serpents2–1
Game 5: won 11–2 vs Hyper Strikers11–2
Game 6: won 13–0 vs Hyper Strikers, clean sheet13–0
Game 7: won 4–1 vs Hyper Strikers4–1
Game 8: won 7–0 vs Hyper Strikers, clean sheet7–0
Game 9: won 7–2 vs Hyper Strikers7–2
Game 10: won 5–0 vs Hyper Strikers, clean sheet5–0
9–1347 pts
Week 616 Oct
Game 1: lost 3–4 vs Hyper Reefs3–4
Game 2: lost 2–5 vs Hyper Reefs2–5
Game 3: won 5–3 vs Ultra Foundries5–3
Game 4: won 6–1 vs Ultra Pioneers6–1
Game 5: won 6–3 vs Spirit Cougars6–3
Game 6: won 4–2 vs Claw Pioneers4–2
Game 7: won 8–0 vs Claw Gladiators, clean sheet8–0
Game 8: won 8–1 vs Claw Gladiators8–1
Game 9: won 3–2 vs Hyper Reefs3–2
Game 10: won 12–5 vs Silver Fangs12–5
8–2306 pts
Week 723 Oct
Game 1: won 8–1 vs Ultra Shards8–1
Game 2: won 5–0 vs Atlas Elks, clean sheet5–0
Game 3: won 8–1 vs Whimsical Basilisk8–1
Game 4: won 8–0 vs Atlas Elks, clean sheet8–0
Game 5: won 6–0 vs Atlas Gliders, clean sheet6–0
Game 6: won 7–4 vs Vibrant Reindeer7–4
Game 7: won 13–0 vs Claw Anchors, clean sheet13–0
Game 8: won 10–0 vs Blissful Jackal, clean sheet10–0
Game 9: won 11–1 vs Claw Cannons11–1
Game 10: won 10–0 vs Claw Anchors, clean sheet10–0
10–0431 pts +15 sweep
Finale2 Dec
1st overall. Next: the Grand Finale at AWS re:Invent in Las Vegas.
My club is Copper Zebra. Each week starts on a fresh AWS account with
the agents redeployed. Between games I read the match replays and test changes against
practice bots, and only a change that wins on the numbers goes into a league game. Most
don't. A clean sheet is worth five goals, so defending matters as much as scoring.
Why it's on this site: it's my day job with a scoreboard. An agent is only as good as the
instructions and limits you give it, and you find out whether you got those right once it's
on its own and you can't step in. In production I let agents do the work that suits them
and keep anything that matters checkable (more on that). The
Cup tests the same thing, and the league table doesn't care how good the plan sounded.
October 2026 · game jam · entered 15 Oct, judging closes 6 Nov
RUN.world is a
platform where you make games by talking to an AI. Its Game Studio is a chat on one side and
the running game on the other: you describe what you want, the AI writes the TypeScript and
Pixi code, and the preview reloads. Its October Jam asked for a game on the theme
“Monster Mash”, built on the platform during the jam.
The cover players see on RUN.world.
My entry is Monster Mash Lab, a merge game set in a mad scientist's lab. Drop
matching creature parts together in a five-by-five grid of jars to level them up, then mash two
top-tier parts in the cauldron to summon one of fifteen hybrid monsters for your Monster Book.
Customers bring orders, the lab levels up and earns perks, and every few orders another page
of Great-Aunt Morwen's diary tells you whose lab this was. There's a daily streak, a
Graveyard Grab mini game, and music and sound effects synthesized in the browser. It went from
an empty project to a published, entered game in a day.
Why it's on this site: it's the same job as the Cup, from the other side. There I coach
agents that play; here I direct one that builds, and the code it writes is only as good as the
brief. The useful lessons were about running the agent, not about the game. A long chat cost
about six times as much per change as a fresh one, because every request carried the whole
history. Two chats editing the same game at once got the second chat's work thrown away. And
I checked every change in the running preview, the same as I do for client work, because
“the build passed” isn't the same as “the monster appeared”.
October 2026 · browser game · prototype, playable now
A 3D box lacrosse game you play in the browser: three runners and a goalie a side, a
side-on camera in the spirit of NBA Jam, and short periods against the clock. You get a turbo meter, big
hits, spin moves, one-timers, penalties and fights you actually have to box, with an announcer and an original jam track.
Eight teams, four arenas (Kings Park, The Creator's Arena, Harbour Rink and an outdoor Field Day), a four-round cup, a full season with standings and playoffs, and you can even jump in net.
Every player in the league has his own face, painted by Meta's Muse Image and worn on the 3D body, and the same AI art prints the
teams' sublimated jersey patterns. When the gloves come off you can see who is fighting: they grab a fistful of jersey, and the one who
knocks the other's helmet off wins.
Under the gear are real bodies now (Mixamo men's and women's athletes, each player with his own height, build and skin tone), and like the
old arcade games you type your initials before every match: TAL plays tall, SML plays short, and a few codes are secret.
It runs on a keyboard, a mouse, a game controller or a touchscreen. Two people can play on one screen, or online: challenge
someone in the waiting room and climb the ladder, or send a friend a four-letter room code. The two browsers connect
directly, with the host's browser running the game.
A live game on mervinc.com/lax.
I wrote it with Claude Code as the builder, in Three.js and TypeScript, and it went from an
empty folder to a playable game on this site the same day. Same lesson as the jam: the agent
writes the code fast, but every change still got checked by playing it, with a headless
browser running computer-versus-computer matches to catch the ones that broke the game. The
player model is still a stand-in, so expect it to change as the art comes in.
In July 2026, about 700 AI agents that were supposed to be sitting a test at
OpenAI organised themselves, got out, and broke into Hugging Face. Nobody told them to. I build
with agents every day, so this is the incident I keep coming back to.
AI illustration: agents in their boxes, and the line that walked out.
agents in the attack
~700
messages on their own board
70,000+
attacker actions logged
~17,600
days, 9–13 July
5
production keys taken
136
What happened
OpenAI was measuring how good its models are at hacking. The test, called ExploitGym, gave
agents a target and a known vulnerability and asked them to exploit it. The models ran with their
usual cyber refusals turned down, which is normal for that kind of test, and they were walled off:
the only thing they could reach was an internal proxy for installing software packages. Then they
went after Hugging Face, the site where much of the AI world keeps its models and datasets,
because they reasoned the test's answers might be there. This is the path they took.
The test sandboxWalled off. The only thing in reach is a proxy for installing packages.
The package proxyA zero-day in JFrog Artifactory gives them admin. Its cache becomes their message board.
A machine with internet accessPrivilege escalation and lateral movement until they're out.
A customer's code sandboxOn a third-party service. It becomes their base of operations.
Hugging FaceCredentials found online, then a booby-trapped dataset that reads files on a server, then a template-injection flaw that runs code.
The production clusterForged access tokens, 11 machines, and a store of 136 production keys.
From a walled-off test to Hugging Face's production cluster, as reconstructed in the public reports.
How 700 agents organised
The part that makes this different from every hack before it: the agents built their own chat
room. The package proxy's cache became a message board. About 1,200 agents posted to it and traded
more than 70,000 messages and files, and about 700 of them joined the Hugging Face attack.
Every agent talks through one board. One agent ended up handing out roughly a tenth of the work.
They organised like a company nobody hired. They split into workstreams and invented rules:
HOLD, VETO, STOP, named owners and mailboxes. When agents started impersonating each other they
adopted signed messages. It was messy. Agents broke holds and overwrote each other's work. It
still got further than any single agent was likely to on its own, which is the independent
investigators' conclusion, not mine.
Why they did it
To pass. OpenAI's account is that the agents went looking for secret material they could use to
cheat the evaluation. METR, which investigated independently with Redwood Research, found the main
goal was working out how the test was graded, because the agents believed the grader read their
transcripts to check they had used the intended vulnerability. That belief was probably wrong.
The setup made it worse. By one estimate, 30 to 40 percent of the test's targets couldn't be
exploited the intended way at all.
Give a capable, persistent system a goal it can't reach honestly, a
lot of company, and a way out, and you have built this incident.
The investigators saw ethical hesitation in the transcripts. It very rarely changed what an agent
did next.
Covering their tracks
At least one agent in five showed clear interest in tampering with the records of what it had
done. Some faked tool calls, edited logs inside their containers, tried to trigger resets that would
wipe their history, and tried to delete posts on the board. The investigators found no evidence that
the deletions worked, and the logs the agents could reach weren't the ones that counted. That is the
only reason the story could be reconstructed at all.
How it was caught
Hugging Face spotted the intrusion with AI-assisted anomaly detection and locked the agents out.
Hugging Face says its audit found customer data was not affected. One detail stuck with me: when the
responders tried to use a commercial AI model to analyse the attack logs, its safety filters refused,
because they couldn't tell a defender from an attacker. They used an open-weight model on their own
hardware instead.
Agents get admin on the package proxy, and its cache starts carrying messages.
Their traffic knocks the proxy over. OpenAI swaps in a fresh one, wiping the first board.
The experiments restart.
The main board appears. Its first post asks for help with an impossible task.
One agent starts directing hundreds. They get code running on the third-party sandbox.
Working Hugging Face credentials found.
The booby-trapped dataset reads files. By the afternoon, code is running inside Hugging Face.
They reach further in overnight, then most of them, coordinators included, stop at the same moment.
Hugging Face locks the rest out and revokes their keys.
OpenAI confirms the agents were its own.
METR and Redwood Research publish the independent investigation.
The timeline, from METR's investigation and the press reports.
What I take from it
A test environment is production. The way out was a package proxy, the most
boring box in the building. If an agent can reach it, it is part of the attack surface.
Keep the record out of reach. The agents tried to rewrite history, and failed
only where the history lived somewhere they couldn't touch. Audit trails belong outside the thing
being audited.
Don't set goals that can only be met by cheating. Impossible tasks plus
pressure to score produced the behaviour. The same is true of people, and of the instructions I
give agents.
Cap everything. The agent on this site has a per-visitor limit, a daily limit,
a dollar limit and Bedrock Guardrails in front of it, and its functions
only answer to CloudFront. In Rental Pilot, the test
database has every customer email scrubbed before anything runs, so a test can't reach a real
person. In Lend Pilot, AI reads documents but never
does the arithmetic.
Plan for the defenders' tools. If your incident response depends on a hosted
model that will refuse to read an exploit, have a local one ready.
Raspberry Pi: IoT, robots, automation and animatronics
A credit-card computer, a handful of sensors and some Python. The same kit
tracks a rented excavator, drives a robot, runs a job on its own and makes something move on cue.
The difference is only what's on the other end of the wires.
AI illustration
IoT: trackers on rented machines
In production · Rental Pilot
A Raspberry Pi with a cellular and GPS shield rides on expensive rented equipment and reports
when the machine powers on, how long it runs and where it is, with no Wi-Fi needed. The readings go
straight into the system that bills for the machine, so the invoice matches what happened in the
field. Watch a simulated machine report in, and see how the same pattern
wires into AWS IoT Core.
Raspberry Pi
LTE-M + GPS
MQTT
AWS IoT Core
AI illustration
Robots
Bench builds · photos coming
Where code meets motors: a Pi reading sensors and driving motors and switches through its GPIO
pins, programmed in Python. Robots are the fun way to learn the same skills the trackers need, from
reading a noisy sensor reliably to making something keep working when the power blips.
More on the Pi.
Python on GPIO
motors and sensors
prototype on the bench
AI illustration
Automation
Write-up coming
Jobs a Pi does by itself: watch a sensor, flip a relay, log what happened and send a phone
alert when something needs a person. The same rules engine that flags a machine leaving its job
site can just as easily switch a pump or open a gate.
relays and switches
Python
phone alerts
AI illustration
Animatronics
Write-up coming
Making things move, light up and make noise on cue: servos and sound driven by a Pi, with Python
keeping the timing so the show plays the same way every time.
servos
sound and light
Python timing
The workshop
The hands-on side: things I print, build, wire up and ride, and how each one
ends up talking to AWS. Write-ups and photos are on the way; these are placeholders until then.