09 · Agentic workflow
How the fleet was run, and how the ship was steered
One human, one orchestrating Claude session that lived for the whole project. In the first phase (2026-08-11 to 09-03), 147 subagents worked in 96 git worktrees. A second phase (09-07 to 10-19) ran the same discipline at a larger scale, for longer, with a wider toolset — covered in its own section below. Throughout, the human never opened an editor, and held the only console, the only SD card, and the only pair of eyes that could be trusted.
The topology
One long-lived session was the orchestrator. It held the project memory, wrote the briefs, merged the branches, and talked to the user. Work was fanned out to background agents, each given a worktree, a branch named for its lever, a brief file committed into docs/research/, and a progress file it had to write before it was allowed to finish. Agents never talked to the user; the orchestrator relayed. Agents never touched mainline; the orchestrator merged after an emulator smoke test and, for anything risky, after a hardware verdict.
The rules the fleet lived by
These accreted from failures (see Difficulties) and were copied into every brief as a "common rules" block:
| Rule | Born from |
|---|---|
| Patch as pure delta. Upstream submodules stay pristine; every change is a patch generated against the full current stack; the README lists the order; a clean-stack round-trip must pass. | Double-carried hunks in the rival-detail patch, twice. |
| Killswitch for everything. Every lever has an ini key defaulting on, and later a row in the bottom-screen DBG tab, so hardware can A/B without a rebuild. | The first hardware rounds could only compare builds. |
Receipt for everything. A lever must print a log line proving it engaged ([rt] mode=ahead, [c3d] anchor=…, texcacheMainMut=0). | A silent audio agent whose run was attributed to the wrong binary. |
| Commit early, progress file first. Agents die on token limits and API overloads; a dead agent with zero commits costs a day. | Traffic, sky3b, P2 and the first bridge-cache agent. |
One emulator, one lock. /tmp/azahar.lock, touched every 30 s while held; wait in turn; never kill another run. | Agents killing each other's Azahar and deleting SD artifacts. |
| Combined-lever screenshot before staging. A race frame with every lever on, captured through the window id, inspected by eye. | The HUD-destroying interaction between trifast and the atlas. |
| Merge ini keys, never overwrite. The user's SD ini holds their 3D and display choices. | A round that shipped with stereo silently off. |
| Never re-apply the patch list onto a patched tree. Reset both submodules first. | A corrupted decomp tree in the final week. |
Steering from outside the code
The user's messages are short, frequent, and almost entirely about direction, evidence and verdicts. Reading all 223 of them in order, the pattern is a captain, not a programmer: set the goal, demand parallelism, bring the physical evidence, say yes or no, and protect the budget.
Setting the goal and refusing the safe path
Demanding parallelism
Being the oracle
The agents could not see the console. The user ran the SD-card loop dozens of times ("SD in", "SD returned", "SD card is back in" appear 36 times) and reported what a human sees: shapes, colours, hitches, feel. Those reports were the only ground truth for the visual bugs and for every fps number on the Road to 60 page.
Pushing back on the machine
Guarding the budget and the record
Working through the night
Twice (08-20 and 08-21) the user went to sleep with the instruction to keep working. Each night session ran a fleet against a frozen mainline, wrote a morning-status document, and left a candidate branch to promote after the user's eyeball. The 08-20 candidate was promoted with "sure, promote it. i dont see any large regressions. probably better"; the 08-21 night produced the S7 interpreter memo and the async frame-mirror copy that took the menus from 15 to 20 fps in the emulator.
What the orchestrator did that an agent could not
- Held the memory. A single project memory file (591 lines by the end) was rewritten at every milestone with state, gotchas, and the next step, and three RESUME documents were committed at every credit pause. Every context compaction and every new session restarted from it.
- Wrote the briefs. Each lever got a brief with the hypothesis, the file locations, the measurement protocol, the killswitch name, the receipt format and the report format. The briefs are in the repository.
- Finished what agents dropped. The texture-cache crash fix, the viewport fix and the stereo anchor were completed by the orchestrator after the assigned agents died on API overloads.
- Ran the hardware loop. Build, package (.3dsx + .cia + test plan), copy to SD, eject, read the log back, compute the fps distribution, decide merge or revert.
The second phase: longer, wider, with an adversary
The first phase was a 24-day sprint to a single target. The second (09-07 to 10-19, 43 calendar days) built a different renderer, a second console tier and a second game mode on top of a build that already worked, which changed the shape of the work more than its length: many more parallel workstreams, each still in its own worktree with its own brief and progress file, left behind 199 internal progress notes by the end, far more than the first phase's smaller written record. The rules in the table above (patch as pure delta, killswitch for everything, receipt for everything, commit early) held and were leaned on harder, because more of the work now ran unattended for longer between human checks.
Three things were added to the loop that the first phase did not have:
- A network link to the console. The first phase's only path to hardware was ejecting the SD card, copying files, and reinserting it. The second phase added a lightweight network link (a homebrew-loader push plus a small device-link tool) so a build could reach a development console directly over Wi-Fi, cutting the time between a build finishing and a human seeing it run.
- An emulator replay lane for proof without hardware. A record-and-replay harness captured exact input sequences and per-frame state hashes, so a new render path could be checked frame-for-frame against the old one — bit-identical pixels, identical game-state hashes — without needing the console or a human at all. This is what let dozens of rendering changes be verified as behaviourally identical before ever being staged for a hardware round, the same role the emulator played in the first phase but run far more often and more mechanically.
- A standing adversarial review. After a window where many parallel branches landed in quick succession, a dedicated review pass audited the merged result for what the individual branches' own testing could not catch — not new bugs in any one branch, but damage from combining them: two agents measuring the same thing with code that quietly disagreed, a comparison tool whose flag got pasted out of existence, a safety wait that had become unreachable. Findings were ranked by severity and marked confirmed or suspected, and the practice repeated after later merge windows.
What stayed the same: one orchestrating session, one human approving every merge into the branch that builds toward release, and no agent ever touching the only SD card or the only console directly.
Where the workflow was pioneering
Multi-agent coding is usually demonstrated on web apps with a test suite. This project ran it against a target the agents could not observe, with a human-in-the-loop measurement cycle measured in SD-card round trips, on a codebase made of three repositories (decomp, runtime, port) with a patch stack instead of a fork. The Frontier page names the devices that made that work transferable. What this page adds: the second phase runs the same argument at a larger scale — the same devices and discipline, held across more than twice the calendar time, more console configurations, and a ground-up renderer rewrite.