The plan: go wide without making a mess
Seething Sheep’s first world, the Meadows, was already playable. The other three worlds (the Wildwood, Frostfell and Wolfcrag) had their art finished but not their code. That is a lot of separate pieces: new sheep, new wolves, new hazards, four bosses, and the levels themselves.
The plan, written the day before, split the work into jobs that could run side by side, each handled by its own agent working in its own copy of the code. It also fixed what went wrong the last time I tried a parallel run, on 4 and 5 October: agents kept editing the same shared files and getting in each other’s way. So a short first step turned those shared tables into one file per item. A new sheep, wolf, card or hazard became a new file, not an edit to something everyone else was editing.
My standing rule did the rest. Daytime is for building and quick checks. The slow, heavy checks (the level solver, the bots, the full test suite) run overnight.
What got built
By the end of the evening there were 40 jobs merged, with the quick set of tests passing, 731 out of 731. A merge agent ran those same quick checks on the combined code before each branch landed. Here is what the jobs added:
- Sheep: Sponge, Crystal, Pilot, and the Sheep in Wolf’s Clothing.
- Face cards: Queen, King and Ace.
- Hazards and world pieces: ropes and rope bridges, rolling logs and conveyors, snow and thin ice, avalanches, the wolves’ return-fire catapults, caged lambs, lava and iron.
- Wolves: the Woodcutter, the Frost Wolf and the Iron Wolf.
- Bosses: the Lumberjack Wolf, the Ice Witch, the Siege Engineer and the Pack King, with the ending.
- The Wildwood itself: 15 levels, plus a Hard level, a Puzzle and Timber Rush.
- The rest: new sound effects, achievements, accessibility options, save export, and a quicker load for each world.
The Wildwood levels went in with provisional star lines, estimated from the agent’s own winning line. Replacing those with solver-backed numbers was a job for the night.
Every agent had to report back
Each agent finished with the same written report, under the same headings, and they were all gathered into one findings file. “The lab” is my shared game engine and tooling.
### Bugs (game)
### Bugs (lab)
### Improvements (game)
### Improvements (lab)
### Decisions the job made
### Questions for Jonathan
A good report is boring and specific. One engine job, for example, kept the physics version number where it was, because raising it would have broken the stored test replays of the Meadows levels, and it wrote down why. The “decisions” heading is for every time an agent had to interpret the design: it says what it chose, and why.
Then, at 00:15, the internet dropped
The night run starts at midnight. Here is how the next few hours went:
- Evening, 6 OctoberAll 40 jobs merged.
- 00:00The night run starts: solver, bots, the full test suite.
- 00:15The internet drops.
- 00:22It comes back. The night run’s session had already exited, 22 minutes in, with no report.
- 00:33The night run is started again, and runs until about 06:25.
- By 08:00A wave of fixes for what the night found is merged.
In those seven minutes the merge queue lost four merges and the end-to-end smoke check. The orchestrating agent merged three of the branches by hand, running the quick tests and the build after each. One of them still clashed on a generated file, so it was regenerated.
A workflow agent that dies on an API error leaves its step undone with no retry.
Lab lesson, from the findings file the same night. The fix: back off and try again on a network failure before giving up.
The night’s verdict: red
The night run’s job is to be unkind to the day’s work. Its report began: “The game is broadly sound, but no level could be safely changed tonight and the checking tools are broken.” The full test suite ran 1,979 tests: 1,974 passed, 4 failed (all already known) and 1 was skipped. The three findings that mattered most:
- A freeze. On one Wildwood level, a Blaze burst that broke a rope made the game throw an error and freeze with no message until you restarted. It hit about one shot in 200.
- Two tabs, one save. With the game open in two tabs, one tab silently overwrote the other’s progress.
- A hard start. The solver could not find a dependable winning line for the first Wildwood level, so the first Wildwood level may be much harder for players than intended.
The fix wave that followed put all eight fixes in by 08:00: the rope crash, the two-tab save, the broken checking tools, solved star lines for three levels, both Puzzle levels redesigned, two of the early Wildwood levels made more forgiving, and a handful of small interface fixes.
What isn’t there yet
Frostfell and Wolfcrag have their systems built and their art finished, but no levels. Those 36 levels were always going to wait for my playtest of the Wildwood. On 7 October I paused Worlds 3 and 4, so the focus is polishing the Meadows and the Wildwood.
And then the name changed
Also on 7 October, DropForge stopped being the name. The games now live at Hillmade Games, at games.hillmade.uk, and dropforge.ai sends visitors to the new address.
I renamed it from DropForge to Hillmade Games because I want a cohesive identity. I am a builder, a maker, a creative, a father, a husband, and it's all rolled into one. Hillmade represents me, my work, and my family legacy. DropForge conversely is a cool sounding name for someone who was too happy to purchase a cheap domain. So welcome to Hillmade Games.
What I took from it
What I took from that night was, AI is a fantastic tool when used in the right way, and the right way doesn't mean having a fancy claude.md file, it doesn't mean having twelve hundred plugins, it doesn't mean having years of coding experience to know exactly how things should work. Instead, it's about having a vision, and being able to work collaboratively with your tools to set clear, auditable, repeatable chunks of work, even where they can be parallelised, so that the agents know exactly what they need to do. Seeing 40 agents move beyond what my previous agent spent 2 days going round in circles thinking about doing was a great relief.