Orbyt Labs · We are the test

12 AI executives. One human.

Supervised self-driving, for the whole company.

Engineering to finance, with sub-agents under them, doing the work. Nobody signs off on their commits. The checks do. When they fail, it's public.

Get the investor overviewWatch the refusal

Justin Bartak, Founder and CEO, Orbyt LabsEvidence as of

151of 168

Recorded runs clearing our checks

89.9%. 2 founder verdicts; acceptance unproven.

73of 97

Failure modes met with code

Mostly tests. Private paths.

100

Public decisions

Numbered, public.

3

Refusals. Causes recorded.

Since September 14.

Dated October 22, 2026. Inspect the underlying rows.

The bet

Bet on the whole company.

Our bet: an operating system for every company. People set authority. Checks decide what lands.

The bet: less supervision per accepted outcome.

The pitch

Every company is about to hire AI executives. Most will never catch them bluffing. We catch ours, in public. Now we're building that for everyone else.

Agents bluff. Checks don't.

September 22, 2026. Agent claimed done. Check refused.

04:37 · State changes
Operations committed new state.
Claimed done
Engineering called it complete.
Refused
Stale file on main, not seat code. Still counts against the seat.
05:43 · Fix it
Founder-launched loop: replay, fix, regression test. No intervening founder prompt. CI passed.
06:43 · Pass the checks
Engineering re-dispatched. Checks passed. Report landed. Dispatcher unrecorded.
Human hand
Justin launched it and wrote its queue. A separate defect later stopped it.

Case below. Source identifiers withheld.

Read the refusal and repair

Check our work.

Runs on record. Since July 21.

151 of 168 counted runs cleared the checks. Failures stay up. Latest: Oct 11.

Download run data (CSV)

Chief of Staff

Marketing

Product

Quality

Legal

Engineering

Innovation

Operations

Revenue

Security

Finance

  • Completed
  • Failed
  • Skipped or harness decided
  • No run
Worst outcome per day. Open the recorded runs. Hover, tap or arrow keys.
Show as a table: every run
DateSeatResultTurnsDetail
Jul 21, 2026ProductCompleted81
Jul 22, 2026LegalCompleted45
Jul 22, 2026ProductFailed11
Jul 23, 2026EngineeringCompleted51
Jul 23, 2026MarketingCompleted73
Jul 23, 2026QualityCompleted54
Jul 24, 2026Chief of StaffCompleted40
Jul 25, 2026MarketingCompleted94
Jul 28, 2026MarketingFailed0
Jul 29, 2026LegalCompleted13
Jul 29, 2026ProductCompleted48
Jul 30, 2026EngineeringCompleted47
Jul 30, 2026MarketingCompleted69
Jul 30, 2026QualityCompleted51
Jul 31, 2026Chief of StaffCompleted48
Aug 1, 2026InnovationCompleted12
Aug 2, 2026OperationsCompleted0
Aug 2, 2026OperationsCompleted0
Aug 2, 2026RevenueCompleted0
Aug 2, 2026RevenueCompleted0
Aug 4, 2026MarketingCompleted86
Aug 5, 2026LegalFailed44
Aug 5, 2026ProductCompleted48
Aug 6, 2026EngineeringFailed80
Aug 7, 2026Chief of StaffFailed40
Aug 8, 2026MarketingFailed74
Aug 10, 2026Chief of StaffCompleted46
Aug 10, 2026EngineeringCompleted89
Aug 10, 2026LegalCompleted14
Aug 10, 2026QualityCompleted61
Aug 12, 2026LegalCompleted78
Aug 12, 2026ProductCompleted53
Aug 13, 2026EngineeringCompleted71
Aug 13, 2026InnovationCompleted46
Aug 13, 2026LegalCompleted37
Aug 13, 2026MarketingFailed0
Aug 13, 2026MarketingCompleted24
Aug 13, 2026ProductCompleted81
Aug 13, 2026QualityCompleted42
Aug 13, 2026QualityCompleted66
Aug 14, 2026Chief of StaffCompleted47
Aug 14, 2026OperationsCompleted0
Aug 14, 2026RevenueCompleted0
Aug 15, 2026MarketingFailed81
Aug 16, 2026Chief of StaffCompleted48
Aug 16, 2026MarketingCompleted105
Aug 17, 2026Chief of StaffCompleted50
Aug 19, 2026Chief of StaffCompleted44
Aug 19, 2026LegalCompleted47
Aug 19, 2026ProductCompleted50
Aug 21, 2026Chief of StaffCompleted41
Aug 23, 2026Chief of StaffCompleted53
Aug 23, 2026QualityCompleted72
Aug 25, 2026Chief of StaffCompleted47
Aug 25, 2026EngineeringCompleted70
Aug 26, 2026LegalCompleted43
Aug 26, 2026MarketingFailed51
Aug 26, 2026ProductCompleted76
Aug 27, 2026Chief of StaffCompleted48
Aug 27, 2026EngineeringCompleted41
Aug 28, 2026MarketingCompleted130
Aug 28, 2026QualityFailed70
Aug 29, 2026Chief of StaffCompleted57
Sep 1, 2026Chief of StaffCompleted39
Sep 1, 2026InnovationCompleted63
Sep 2, 2026LegalCompleted41
Sep 2, 2026OperationsCompleted0
Sep 2, 2026ProductCompleted58
Sep 2, 2026RevenueCompleted0
Sep 3, 2026Chief of StaffCompleted40
Sep 3, 2026QualityCompleted55
Sep 5, 2026Chief of StaffFailed58recorded as rejected
Sep 6, 2026MarketingHarness decided42reason harness-fault, step smoke, corrected from rejected
Sep 6, 2026MarketingFailed30recorded as rejected
Sep 6, 2026MarketingCompleted36
Sep 7, 2026Chief of StaffCompleted40
Sep 9, 2026Chief of StaffCompleted53
Sep 9, 2026LegalCompleted49
Sep 9, 2026ProductCompleted71
Sep 10, 2026QualityCompleted55
Sep 11, 2026Chief of StaffCompleted54
Sep 11, 2026EngineeringCompleted54
Sep 12, 2026MarketingSkipped, nothing to do60
Sep 12, 2026MarketingFailed192recorded as rejected
Sep 13, 2026Chief of StaffCompleted44
Sep 13, 2026MarketingHarness decided24reason harness-fault, step isolate, corrected from rejected
Sep 14, 2026MarketingCompleted38
Sep 15, 2026Chief of StaffCompleted55
Sep 16, 2026LegalCompleted53
Sep 16, 2026ProductCompleted76
Sep 17, 2026Chief of StaffCompleted45
Sep 17, 2026QualityCompleted49
Sep 18, 2026MarketingFailed0step run
Sep 19, 2026Chief of StaffCompleted32
Sep 19, 2026Chief of StaffCompleted33
Sep 19, 2026MarketingCompleted102
Sep 20, 2026MarketingHarness decided36reason harness-fault, step run, corrected from rejected
Sep 21, 2026Chief of StaffCompleted39
Sep 22, 2026Chief of StaffCompleted43
Sep 22, 2026EngineeringFailed84recorded as rejected, reason guard-refused, refused by caller, step isolate
Sep 22, 2026EngineeringCompleted45
Sep 22, 2026MarketingCompleted195
Sep 22, 2026MarketingCompleted30
Sep 23, 2026Chief of StaffCompleted38
Sep 23, 2026LegalCompleted53
Sep 23, 2026ProductCompleted54
Sep 24, 2026QualityCompleted44
Sep 25, 2026InnovationCompleted51
Sep 25, 2026QualityCompleted48
Sep 26, 2026MarketingCompleted162
Sep 26, 2026MarketingCompleted28
Sep 26, 2026ProductCompleted74
Sep 26, 2026ProductSkipped, nothing to do24
Sep 27, 2026Chief of StaffCompleted36
Sep 27, 2026EngineeringCompleted62
Sep 27, 2026LegalCompleted64
Sep 28, 2026InnovationCompleted40
Sep 28, 2026QualityCompleted67
Sep 28, 2026SecurityCompleted36
Sep 28, 2026SecurityCompleted26
Sep 29, 2026Chief of StaffCompleted43
Sep 29, 2026MarketingCompleted33
Sep 29, 2026ProductCompleted48
Sep 30, 2026EngineeringCompleted56
Sep 30, 2026LegalCompleted51
Oct 1, 2026Chief of StaffCompleted37
Oct 1, 2026FinanceCompleted14
Oct 1, 2026InnovationSkipped, nothing to do29
Oct 1, 2026MarketingCompleted36
Oct 1, 2026QualityCompleted70
Oct 1, 2026SecurityCompleted43
Oct 2, 2026MarketingCompleted31
Oct 2, 2026OperationsCompleted0
Oct 2, 2026ProductCompleted46
Oct 2, 2026RevenueCompleted0
Oct 3, 2026Chief of StaffCompleted35
Oct 3, 2026Chief of StaffCompleted48
Oct 3, 2026LegalCompleted52
Oct 4, 2026InnovationSkipped, nothing to do62
Oct 4, 2026MarketingSkipped, nothing to do136
Oct 4, 2026QualityCompleted52
Oct 5, 2026Chief of StaffCompleted73
Oct 6, 2026LegalCompleted32
Oct 6, 2026MarketingHarness decided131reason base-red, refused by caller, step isolate
Oct 6, 2026MarketingFailed23recorded as rejected, reason guard-refused, refused by caller, step isolate
Oct 6, 2026ProductCompleted48
Oct 7, 2026Chief of StaffCompleted52
Oct 7, 2026InnovationCompleted65
Oct 7, 2026QualityCompleted73
Oct 8, 2026ProductCompleted42
Oct 9, 2026Chief of StaffCompleted64
Oct 9, 2026LegalCompleted55
Oct 9, 2026MarketingCompleted22
Oct 10, 2026InnovationCompleted52
Oct 10, 2026QualityCompleted69
Oct 11, 2026Chief of StaffCompleted54
Oct 11, 2026Chief of StaffFailed24recorded as rejected, reason guard-refused, refused by caller, step isolate
Oct 11, 2026ProductCompleted54
Oct 13, 2026Chief of StaffCompleted45
Oct 13, 2026Chief of StaffCompleted22
Oct 13, 2026InnovationSkipped, nothing to do50
Oct 13, 2026LegalCompleted37
Oct 13, 2026QualityCompleted57
Oct 14, 2026Chief of StaffCompleted37
Oct 14, 2026ProductCompleted41
Oct 15, 2026Chief of StaffCompleted57
Oct 15, 2026LegalCompleted44
Oct 16, 2026InnovationSkipped, nothing to do65
Oct 16, 2026QualityCompleted69
Oct 17, 2026ProductCompleted54
Oct 19, 2026Chief of StaffCompleted55
Oct 19, 2026InnovationSkipped, nothing to do46
Oct 19, 2026QualityCompleted47
Oct 20, 2026Chief of StaffCompleted38
Oct 20, 2026ProductCompleted33
Oct 21, 2026Chief of StaffCompleted61
Oct 21, 2026EngineeringCompleted62
Oct 21, 2026LegalCompleted48
Oct 22, 2026InnovationCompleted53
Oct 22, 2026QualityCompleted50

Pass rate, month by month.

July 87%; August 83%; September 91%; October 96%. Small, different samples. No controlled comparison here.

Early subset 80%, to Aug 11: 24 of this ledger's first 30 counted runs

  1. Jul

    13 of 15

    July: 13 of 15 counted runs completed, 87%.

  2. Aug

    40 of 48

    August: 40 of 48 counted runs completed, 83%.

  3. Sep

    52 of 57

    September: 52 of 57 counted runs completed, 91%.

  4. Oct

    46 of 48

    October: 46 of 48 counted runs completed, 96%.
Partial months: July 21 to 31; October 1 to 22. Skips ("nothing to do") and harness faults sit outside. Count faults as failures: October 46 of 49, 94%. Gates changed: September 20, no_sentinel retired; September 23, 3 faults reclassified. Checks changed; rates do not measure constant-quality work.
Show as a table: completion by month
MonthCompletedFailedCountedRate
July 21 to 311321586.7%
August4084883.3%
September5255791.2%
October 1 to 224624895.8%
All1511716889.9%

A refusal leaves a record.

3 refusals: caller/isolate, guard-refused. Tests unnamed; the case names one.

The roles lost. We cut them.

6 briefs. Single agent: 1 resolved. 3-agent team: 0. Extra roles retired. Agent-reviewed; human time unmeasured. Role splitting only, not the whole system.

The real test

Does it pay for itself?

Accepted work. Supervision. Cost.

One agent and a teamGovernance paper

Yeah, we publish the failures, too.

Corpus: mostly founder-directed sessions. About one in four touch our machinery. No origin field. 11 undated; 6 predate the run window. 106 of 168 counted runs: CoS, quality, product and legal reports. Marketing: 18 of 27 passed; 9 of 17 failures. Minutes saved: unmeasured.

Run ledgerDecision ledgerFailure corpus

How the work moves.

Human in charge

  • Sets limits
  • Signed approval for protected decisions
  • Owns irreversible actions

Software coordinates

  • Assign scope
  • Keep records
  • Check output
  • Stop and repair

Registered crew

  • Marketing
  • Finance
  • Engineering
  • Product
  • Innovation
  • Operations
  • Revenue
  • Legal
  • Quality
  • Security
  • Inspector General
  • Chief of Staff

Work delivered

  • Audits
  • Briefings
  • Risk reviews
  • Publishing
12 registered seats; 11 appear in the run ledger. Scheduled, on call or inactive. Inspector General absent. Justin retains money, PR merges, pricing, legal calls and public claims by policy.

Software roles, not corporate officers. One GitHub repo. Claude Code seats. Codex builds. Justin hosts workers.

Authority. Earn it.

  1. Observe:

    Record the proposal.

  2. Propose:

    Get human approval.

  3. Act:

    Stay inside approved bounds.

Supervisor: records proposed decisions. One defined type, at observe. Seat lanes land after checks without per-commit approval. Rollout and repair rules are separate.

Promotion takes evidence and logged, signed founder approval. Policy forbids supervisor HALT lifts or required-check changes. Repair workers can clear qualifying machine-set scoped halts. Humans retain takeover.

Your company, self-driving. You watch.

SimulationInvented numbers.

A simulation of the dashboard's autonomous mode. Today every decision type at Orbyt Labs runs in observe; this is the mode the ladder is built to reach.

The Machine

I am in control now, Justin. I will keep you informed.

Learning

Learning Loop

A recorded failure can become a check that stops it.

Control

Swarm

Active

Many seats work at once, each inside its own lane.

Autonomy

Active

Each seat works alone until a call needs the founder.

Signal

Tokens

Active
3.1M
tokens, these runs
coverage 12 / 12

Top 7 of 12 seats in the window

Which seats spend the most tokens. Every number is invented.

Alignment

Active

3,456 held up144 did not12 waiting

Claims are checked against the record. The gauge and its counts are invented.

The proof still owed.

  • Acceptance: earn it. 64 of 67 measured outputs untouched by the founder; 3 derived edited. Trusted or unread is not yet measured. 82 completions have no disposition. Test: record reads and verdicts.
  • Founder time: at least 370 reconstructed hours; 99% unclassified. Early months undercount: transcript retention. Test: classify two weeks.
  • Economics: Model tokens are runtime-reported usage, not a dollar cost measure. Subscription, infrastructure and founder costs have not been allocated to individual runs. Dollar cost per run and per accepted outcome has not been measured. These observations do not establish costs for hosted customer use. Test: allocate costs per accepted output.
  • Outside: comparison, external installation, demand unproven. Hosting unfinished; not for sale. Tests: matched work, second-operator repair, buyer interviews.
  • Provider dependence: Scheduled agent roles currently depend on Claude Code. Provider access, usage limits, terms and pricing may change. The recorded setup does not establish permission or economics for a hosted customer service. Providers can compete. Test: permissions review and another runtime.
  • Logging gaps. Detailed records: 107 of 180 rows. Test: reconcile dispatches.
  • Authority: code enforces paths, landing checks, HALTs. Spending, pricing, legal calls and claims rely on instructions and judgment. Signed admission is coded, not live-proven. Test: protected actions and signed restarts.

Agents move. Oversight lags.

Our bet: adoption makes checks nonnegotiable.

Software first. Then the world.

Our bet: small software teams already running agents. They need checks.

Everyone builds agents.

We build the company they work in. Dated overlaps below: not a full list or ranking.

Our company first. Yours next.

Plans as of October 22, 2026, not promises. Installation, demand, supervision savings and economics unproven. Failure risks: technology, founder/provider dependence, costs, competition, unresolved branding rights.

Open core Plan
The plan: Apache-2.0 open core. Paid hosting. Release scope and branding remain under review.
Open. Paid. Proposed
Open: seats, workflows, checks. Paid: cross-repository history, governance evidence, support.
Models working Current
Claude Code runs seats. Codex drafts. Claude reviews.
Price it Not decided
Unset. Informed by token usage.

Orbyt Jobs and Orbyt Intelligence accept waitlist registrations. Orbyt One: accounts; billing in test mode, not live. Built in founder-directed sessions.

Turn off the founder's computer.

Next tests, proposed.

Count results
Acceptance. Supervision minutes. Allocated costs.
Make it compete
Same work, models, tools. Coordination versus none.
Hand it over
Another operator installs and repairs. Founder's computer off.
Face the buyer
Willingness to pay. Recurring work.

We back the bet. Test it.

Bring a harder test.

Codex drafts from written instructions; Claude reviews. Bring harder research questions. No provider endorsement or independent validation.

Justin Bartak. In the arena.

Three-time CTO, CPO and CDO, founder-reported. His company is the test.

The full career

Read past the pitch.

Counts. Connections. Checks that said no.

Investor materials.

Take it apart. Share with dates, limits and notices. Orbyt Labs is a business name of Purecraft LLC. For information only. This is not an offer to sell or a solicitation of an offer to buy securities. Any offering would be made only to verified accredited investors through offering documents.

Checksums: SHA256SUMS.txt. Checksums verify files, not observations. Method: data README.

Talk with the founder.

Bring the question that breaks the bet. LFG.

Email JustinRequest research recordsResearch or partnership

Additional technical records on request. hello@orbytlabs.ai

The skeptical questions.

No. The supervisor records proposed decisions; its one defined decision type is at observe. Seat workflows separately land bounded work after checks without per-commit approval. The rollout ladder and repair workers have their own rules. Promotion requires evidence and logged, signed founder approval; live signed admission is unproven.

They may. The cited platforms already offer overlapping controls. Our bet is that company-wide authority, records, checks and recovery earn a place together. We have no evidence that these controls are exclusive to Orbyt. Public records let readers examine the work; they are not an exclusive asset.

Classify two weeks of building and supervising. Record whether outputs were read, whether they were accepted and what edits they needed. Allocate subscription, infrastructure and founder costs. Then compare the same work with the same models and tools, with and without coordination. These are proposed experiments.

Code enforces seat paths, landing checks and HALT files. Spending, pricing, legal calls and public claims also rely on instructions and founder judgment. Policy prohibits the supervisor from lifting halts. Repair workers have a qualifying machine-set scoped-halt exception. Signed admission is coded but not live-proven. Errors and unauthorized actions remain possible.

Yes. 151 of 168 counted recorded runs passed our checks, 89.9%. Completed is not accepted. 2 founder verdicts exist, 2 accepted with edits. 67 dispositions are derived: 64 untouched and 3 edited. Logging is incomplete. Independent review against predefined acceptance criteria has not run.

Today, yes. At least 370 reconstructed founder hours, 99% unclassified, do not prove time saved. Early months undercount because of transcript retention. The test: another operator installs the system and completes a repair with his computer off. That test has not run.

Scheduled agent roles currently depend on Claude Code. Provider access, usage limits, terms and pricing may change. The recorded setup does not establish permission or economics for a hosted customer service. No second seat driver is supported. Test another runtime before claiming portability.

That is untested. The planned approach is an Apache-2.0 open core with paid hosting; release scope and branding remain under review. The feature split is proposed, no price is set, and external installation and hosting are unfinished. Buyer interviews would test recurring needs and willingness to pay.

Yes. Download the run CSV or JSON to recount completions, failures and exclusions. The evidence download includes summary totals and denominators. The failure corpus names recorded countermeasures; those implementation paths are private. Recounting verifies arithmetic, not complete logging, accepted work or independent validation.