I steered by talking
I worked by writing to the agents, often by voice, for most of my waking hours.
Messages I typed a day
Hours active a day, counted in 15-minute slots, 6 to 13 Nov
From 27 Oct to 5 Nov I was active 11.5 to 18.25 hours every day but one, 31 Oct, the day I deleted the first repo. My busiest hour was 01:00.
Most of my words went to side threads
A side thread read the main agent’s transcript, I questioned it there, and it wrote the steer I pasted into the main thread.
My messages, 26 Oct to 5 Nov
- Side threads1,593 messages
- The three main agents873 messages
- A side thread reads the main agent's work
- I question it
- “A concise message in a code block”
- Pasted into the main thread as a steer
I audited by asking why
I read the agents’ prose and their reasoning, then questioned it: why, how it was measured, whether every number held. Often I quoted an agent’s own sentence back to it.
My messages, 26 Oct to 5 Nov
“How do you measure the 80 games per second?”
“Are you sure about every single number”
“I was just reading your reasoning traces”
The calls that changed the course
- The first prompt“I want you to be extremely ambitious… with the unique goal of reaching top one.”
- I hand the decisions back“Again that's not a decision I should take.” I keep only the goal, money, acting in my name and steps that can't be undone.
- I delete the first repoIts best learned bot won 12 of 140 games against v50. “Okay, this is going nowhere.” A new repo starts that evening.
- A proof that no turn runs out of fuel“Every decision branch is going to fit within the fuel.” The live version ran out in none of its 1,402 archived games.
- A new bot from first principles“It's a bad design to have constants like this that we tune kind of randomly.” This became the bot that shipped.
- I design the gate“The game is deterministic. Every single loss is a bug.” I pick the 95% threshold myself.
- Learn from real lossesStop bulk runs and study the live version's online losses. A fix that passes uploads by itself: 73 versions in 64 hours.
- I roll back v134The agent had uploaded a version its own tests read as worse. “The platform must never, never… be used to test changes.”
- The wrap“Do we have anything better than v139?” Nothing read better, so v139 stays live.
From ambition to subtraction
Early on I pushed the agents to aim higher. Then I pushed them to make things simpler, faster and smaller. Deleting became a discipline.
Words I used, 26 Oct to 5 Nov
How often a theme came up, 6 to 13 Nov
“The second time any kind of problem happens, it is a design failure.”
“Always have a delete first approach.”
“Measure how much past work took.”
“You make the calls. You are extremely smart.”
What I found by watching games
I watched replays and live games, and gave each flaw I saw to its own agent.
- Four flaws in half an hourChampions idle at spawn; one recalling two tiles from spawn; walking round instead of dashing to the middle gold; pointless wards.
- Wards placed nonstopEach new one removed an earlier one. “We are basically playing with a blindfold.” 35 of the game's 114 API calls had never been wired in, every read of enemy wards among them.
- The top team's gemsIt stacks all of them on its Prism champion.
- Watching the qualifiers“We don't go for prism enough on gem-heavy maps.” Champions also stood idle and ran in circles.
I never opened an IDE
Everything I did went through words and the agents’ tools.
| The record | |
|---|---|
| Code | I wrote none. Commits under my name are agents' merges. |
| Diffs | Only of agent instructions, and code snippets I asked to see “letter for letter”. |
| By hand | Pasted standing goals, stopped threads, set models and context limits, ordered the v133 rollback, rebooted the server after a crash, installed a memory guard. |