One person, the founder, and a company of agents. The goal is a Machine that runs without him. This page is the record of where he is still in it, reconstructed from his session transcripts, and of how little of that time any rule can name. Its other half is autonomous operation.
The question is where people approve, intervene and keep responsibility, and how that share is measured as the agents take on more.
The rule meant to find the founder stopping the Machine mostly found work on the stop button instead. It labelled 46 activity windows as rescues; only two contained a change to a stop file committed by anyone but a bot or the loop. What follows measures his recorded time with the agents and shows how little of it the records can explain.
Why is the human in the loop only for now?
Orbyt Labs exists to test one question: can AI run a product, then a leadership team, then a company? Every rung we climb moves a decision from a person to the Machine. The goal is a Machine that runs on its own, where no error goes unseen: each one is either fixed by the Machine or handed to the founder with a receipt. So this page describes a line that moves. Today a person still decides money, claims, production data and anything that cannot be undone, and still signs the decisions that bind the company. The ledger records where those hours go, week by week, and when a decision moves from a person to the Machine, the decision log records that too.
How will the founder approve from anywhere?
Today he approves at his Mac. Each approval is an SSH signature from his owner key, made with Touch ID, and the Machine reads it through its authority rules: a decision entry, a raise in an agent's autonomy or the removal of a halt requires a commit that key signed. A scoped halt originally written by evaluator or ladder is exempt from that signature requirement; a global halt is not. By decision D-0030, a signature proves custody of the key, not that he was present.
Launch Keys is the product being built so he can approve away from the Mac. In its design he signs what the agents ask for, commits, deploys and payments, from his iPhone, his Mac or his Apple Watch, with keys held in each device's Secure Enclave. Every request and every decision is signed and can be checked offline, denials included. An optional two-person rule is off by default.
Its status is In Development, the same label the homepage shows. No real signed approval has gone through it yet.
Nothing on this page measures Launch Keys. Until a signed approval is recorded through it, the record holds only the signatures made at the Mac.
What does human in the loop mean at Orbyt Labs?
It means a person who can see what he approves. In March the founder wrote, in Human-in-the-Loop Is Not Enough: "If the system is designed to make informed oversight impossible, the human is not a safeguard. The human is an alibi." This page holds his own company to that sentence, and the first thing it owes is what it cannot see.
- The hours are a lower bound. Each window runs from the first to the last prompt, split wherever the session goes quiet for fifteen minutes, and the split is a choice, not a fact about him.
- NOT OBSERVED after . The ledger's newest window ends then, and the reconstruction has not been rerun since.
- Only agent sessions started inside this repository on his Mac are in it; sessions in the console's repository or a working clone are not. Nearly all are Claude Code, a few Codex. Time in the console, on his phone, in email or thinking in the car is not.
- 1,892 of 1,943 windows carry no label for the kind of work, and the rule that labelled most of the rest was counting the wrong thing.
- Untouched is not unread. A report he read and agreed with leaves the same trace as one he never opened.
- The agent side of the same transcripts holds 770.4 summed hours over 9,741 turns, but his windows contain the agents' turns, so no ratio of agent hours to his is printed. It would count the same minutes twice.
Where does the founder's time go?
Into an effort ledger. Every agent session he runs is cut into windows of activity, and each window is a row. A script rebuilds those rows from the transcripts, so they are an estimate, and the ledger keeps them apart from what he declares by hand. Overlapping windows are unioned, never summed, because one person cannot work the same minute twice.
From to , Central time, the ledger holds 1,943 windows reconstructed from transcripts. Unioned, they come to 410.2 hours. He has declared an intervention by hand twice. Everything else is reconstruction.
How much of that time can a rule classify?
Almost none of it. The ledger has five kinds of intervention: setup, review, correction, rescue and maintenance. Three have a written rule that reads commits under the founder's name, which an agent in his session usually makes, during a window and for 30 minutes after it ends. That extension attributes nearby commits to the window; it adds no minutes to the measured time. The rule reads every author that is not a bot or the loop, so it also counts the commits of push-collision-canary, a test identity that leaked into his git settings. A halt commit is a rescue. A commit that edited or threw away a seat's output is a correction. A window whose every commit touches only files under scripts/, __tests__/, .github/, packages/collective/ or .claude/, or files named CLAUDE.md, is maintenance.
Applied to every window through , the ledger's newest, those rules label 51 of 1,943: 46 rescues, 1 correction and 4 maintenance. The other 1,892 windows, 387.4 hours, stay unclassified. Setup and review have no rule at all, because reading a report or deciding something leaves nothing a rule can find.
These counts come from applying the rules to every window through , not from the committed labels file. That file was last derived on , so a count read from it would be stale.
Were the rescues really rescues?
Mostly not. The rule counted work on the stop mechanism as if the founder had used it to stop the Machine. It matches the word halt, halted or halting in a commit subject, a file or folder anywhere in the repository whose name starts with HALT, or a change to the halt workflow, .github/workflows/org-halt.yml. Each of the 39 commits it used as evidence was checked. Two add, edit or remove a stop file under docs/org, the file that actually stops the Machine. None of the other 37 does. Most were work on, or notes about, the stop mechanism: "a resume must prove it removed the halt itself, not that one is gone", "a halt is dated by the commit that placed it, never by its mtime". Building the stop button is not pressing it.
So a rule that looked precise counted the wrong thing, and nothing reported it, because a label with evidence attached reads as a measurement. The narrower rule, a commit that adds, edits or removes a stop file, finds two windows. Commit 66a568b09, "HALT during the domain flip; pulse tests stand down under a global halt", is under push-collision-canary, the leaked test identity. Commit f0c184819, "install the Security seat, dispatch-only under a human HALT (Agent Sentinel, CSO)", is under the founder's name, and it reads as setup, not rescue: it installed a new seat behind a human stop. That reading is judgment, from its subject. An agent in his session made both.
How often the record backs a rescue
Of 46 windows the rule called rescues, 2 held a change to a stop file.
The rule also cited 39 commits as evidence; 2 of them change a stop file.
- The commits under the funnel are counted apart from it. A window can cite several commits as evidence and two windows can cite the same one, so the commits are not a further stage of the windows.
Built from hitl-figures.generated.json, which scripts/research/method-figures/human-in-the-loop.ts writes from the effort ledger (docs/org/effort/ledger.jsonl) and git log. The ledger is rebuilt from session transcripts, so this is a snapshot through its newest window.
Snapshot through Central
The numbers as a table
| Stage | Count |
|---|---|
| Windows the rule called rescues | 46 |
| Of those, windows that held a change to a stop file | 2 |
| Commits the rule cited as evidence | 39 |
| Of those, commits that change a stop file | 2 |
Download the data: human-in-the-loop-rescue-funnel.csv, human-in-the-loop-rescue-funnel.json
The rule sees only commits whose author is not a bot or the loop. A stop pressed in the console is committed by the executor's bot, and no rule here reads it, so the ledger cannot count those rescues either.
What is the recorded time spent on?
Where a window left commits, the page can say what they touched, which is not the same as saying what kind of work it was. Each window takes the area most of its commits' files fall in. 23 percent of the hours had no associated commit by anyone other than a bot or the loop.
Recorded hours by where their commits landed
The tiles for windows that left a commit add up to 315.3 hours, and the website took 165.7 of them.
- The tiles add to 411.5 hours, 1.3 hours more than the 410.2 hours recorded. Each tile counts its own windows with their overlaps removed, so a stretch where windows in two tiles overlap is counted in both.
Built from hitl-figures.generated.json, which scripts/research/method-figures/human-in-the-loop.ts writes from the effort ledger (docs/org/effort/ledger.jsonl) and git log. The ledger is rebuilt from session transcripts, so this is a snapshot through its newest window.
Snapshot through Central
The numbers as a table
| Where the commits landed | Hours | Windows |
|---|---|---|
| No commit by anyone but a bot or the loop | 96.2 | 781 |
| The website | 165.7 | 588 |
| The Machine itself: seats, workflows and the console's package | 55.1 | 159 |
| The instructions, scripts and tests that steer the agents | 30.5 | 127 |
| Files outside those areas | 29.8 | 161 |
| The books | 24.3 | 79 |
| The research pages | 8.6 | 36 |
| The products | 1.3 | 12 |
Download the data: human-in-the-loop-hours-by-area.csv, human-in-the-loop-hours-by-area.json
By the files the commits touched, not by kind of work. The path map is in the generator, scripts/research/method-figures/human-in-the-loop.ts, so a reader can rerun it with another.
Is the founder's share of the work shrinking?
It cannot be told yet. Here is each month, with the share of its hours that left no commit:
Recorded founder hours by month
The share of hours with no associated human commit fell from 26 percent in August to 16 percent in September, then rose to 31 percent in October.
- These hours do not measure the founder's share of all work.
Built from hitl-figures.generated.json, which scripts/research/method-figures/human-in-the-loop.ts writes from the effort ledger (docs/org/effort/ledger.jsonl) and git log. The ledger is rebuilt from session transcripts, so this is a snapshot through its newest window.
Snapshot through Central
The numbers as a table
| Month | Recorded hours | With no associated human commit | Share with no commit | Windows | Coverage |
|---|---|---|---|---|---|
| July | 1.4 | 0.0 | 0 percent | 7 | Partial month |
| August | 76.9 | 19.8 | 26 percent | 465 | Inside the record's span |
| September | 175.4 | 27.7 | 16 percent | 807 | Inside the record's span |
| October | 156.6 | 48.7 | 31 percent | 664 | Partial month |
Download the data: human-in-the-loop-hours-by-month.csv, human-in-the-loop-hours-by-month.json
The no-commit share fell from 26 percent in August to 16 percent in September, then rose to 31 percent in October. A window with no commit cannot tell reading and deciding from idle time, so the record does not say why the share moved, and this page does not either.
Months are UTC. The first month holds only the reconstruction's earliest windows, and the last ends with the ledger's newest one.
Who catches the problems?
A separate record, kept apart from his minutes so it can never inflate them, says who found each issue. It holds 97 rows: 75 found by a monitor, 16 by an agent, 6 by a guard, and none yet by the founder.
Read the monitor count with care: 39 of the rows are the Machine's own workflows failing and 27 were seeded from older records. The record keeps both and leaves them out of any share.
No founder row is not the same as no founder catch. The record was ordered on , its first row was written on November 5, and its earliest finding is dated September 22. What he caught before that lives in prose. The record that counts it going forward is the AI oversight ledger.
What happens to the seats' work after it lands?
The records label most of it untouched. Of 69 disposition records, 67 were derived by a script on : 64 are labelled untouched and 3 edited. Each label comes from the commits, by anyone but a bot or the loop, that touched the seat's output before that seat's next run, or before the derivation for a seat's latest run. They cover 64 distinct seat and commit pairs, not 67, because three runs matched the same commit as another run of the same seat. In one of those pair the run's commit changed no file in the seat's reports folder, so its label describes no output. The other two were declared by hand, both accepted with edits. Untouched is not unread.
What is the difference between Collaborative and Autonomous?
Two positions of one switch, named by the founder on : "I want to change it to Collaborative and Autonomous instead". In the design's words, Collaborative, which it called Human Loop, starts nothing until the owner says go. Autonomous, which it called Without Me, means the Machine picks its next work, builds, gates and ships with nobody present, inside the same grant, fence and valve. In its first version it governs the handoff loop only.
The design for arming the switch was frozen as decision D-0078 on , before the rename. As of the console rendered the switch, and most of the units that would make Autonomous real were not yet built.
This page has not re-verified the build state since . Autonomy itself starts at observe, by decision D-0053 of .
Why recorded commands instead of hand edits?
Because a hand edit leaves no record of who asked for it. The North Star the company works toward says a human touches the Machine only through recorded commands and directives.
So the console never changes the installation directly. It writes a command record, and a separate executor acts on it. Pause and resume shipped on , then configure, which changes one field of one seat. A press claims only that the command was requested; whether it took effect is a second read from a second source, never the executor's own reply.
Every generated workflow carries the command's id, and no operator may set it by hand, because that id is how a run is attributed to the command that asked for it. The stop a human can always reach is the kill switch, and the rules for who may decide what are on AI governance.
Human in the loop questions? Answers.
Related reading
- Autonomous operation
The dated record of the Machine running without a person in the loop.
- AI oversight ledger
The standing record of who caught each issue and who fixed it.
- The kill switch
How the Machine is stopped, with the dated record of its halts.
- Governing an Agent Leadership Team
The paper whose limitations said active human hours had never been measured, the reason the ledger exists.
- A single agent and a team
Its pre-registered human-time clause reads not measured: an agent, not the founder, read the reports.
- What Your AI Approval Button Hides
The founder on binding an approval to the exact action, and testing that a changed action is refused.
- All Orbyt Labs methods
Every method page, each with its dated public record.
Cite this page
Cite this page for the practice and follow its sources for the underlying records.
Bartak, J. (2026). Human in the Loop. Orbyt Labs Research. https://