Ari’s Space*   est. 2026
Skip to content
ari*s space[ a. ]
← Libraryworkflows / 2026-06-13

540 Commits. One Human.

Eleven days of human planning, delegated execution and independent audit: 540 commits, the reported results and the limits that kept the work bounded.

6 parts / Hover to explore. Click to read.

Eleven days of human planning, delegated execution and independent audit: 540 commits, the reported results and the limits that kept the work bounded.

Plan -> bounded execution
                  |
            independent audit
             /     |       |
           pass   fail   escalate
             |     |       |
           accept repair  human

Measure usefulness, not volume.
The reported workflow checked returns before acceptance and retained escalation. Another model's audit is fallible, while more commits or retrieved context alone do not establish better work.
Context Pack (for agents)

In the eleven days behind this June 2026 account, 540 commits landed. I wrote almost none of that code by hand. I planned the work, routed it and held the decisions that needed a human.

My seat had changed. The useful work was deciding what should happen, which model could do it and what evidence would let the next stage proceed.

Read the numbers as one period of work

These were the reported results from that week and a half:

MeasureResult
Commits540 in 11 days; the busiest day had 106.
Overnight buildSix tasks completed and independently audited, in about two hours without my intervention during the run.
Retrieval graphPrecision moved from 54% to 84%; it then became 9.5 times faster in a day.
macOS appSigned, with 165 tests passing.
Memory engine600 tests passing.
Runaway loopAbout £2 spent before a hard cap stopped it.

The numbers describe that work. A commit count alone doesn't tell you whether a product is useful, and a green test count doesn't show every possible failure. The checks still need to fit what was being built.

Give the plan an execution and review path

For the overnight build, I supplied the plan before going to sleep. Six tasks went through execution and independent audit. By morning, the result was there without me having to intervene in that run.

The rule was that a worker couldn't finish a job merely by announcing success. Another model audited the result and returned pass, fail or escalate before it landed.

That reduced the routine checking I had to do myself. It didn't make another model infallible or remove the decisions I still owned.

Choose the seat for the job

The setup used several model roles. Fable held orchestration. Codex did the heavy coding. Kimi handled long video. MiniMax M3 ran looping builds. Local models supported memory. Opus 4.8 was available when a seat needed more capability.

A model earned its place by what it could do in the system. No single one had to be responsible for planning, implementation, recall and every check.

Let a measurement overrule the design

I built a retrieval approach I thought was smarter. Compared with the simpler baseline, it pulled 29% more context and produced worse results.

I cut that approach. More retrieved material had not meant better help. The comparison gave me a reason to change direction that enthusiasm for the design couldn't supply.

The runaway loop taught a related lesson about limits. It spent about £2 before the cap caught it. I didn't have to wake up and notice the bill growing for the stop to happen.

Keep the human responsible for the system

I called the combined system ARI-OS, with Cortex as its memory layer. The progress was in the arrangement around the models: a plan that could be executed, an independent check before acceptance, retained knowledge and a real cost limit.

Pick one part of your own work where those boundaries are unclear. Decide what goes in, who does it and what would prove it worked. Then test whether you can step back from that part without losing sight of the result.

[ comments ]

guests welcome. members show first in the list

no comments yet. start the thread_

1 / 8