RESEARCH NOTE · 29 OCTOBER 2026

Room to wonder.
Evidence to act.

Our pilots already found and acted on a range of improvements. The next challenge is making that creative usefulness repeatable, affordable and safe as autonomy grows.

Should we keep going?

Yes. Unattended engineering is in real use, and research systems generate and test novel hypotheses. Our own evidence supports continuing too. We have not established dependable, open-ended product judgment or profitable idea-to-production automation for DiveScene.

This review separates our observed results, companies’ reported results and our proposed design. No new scheduler, registry service or deployment is activated.

Discovery demonstrated

Four pilots produced different useful source outcomes. General repeatability and benefit in use remain open measurements.

Imagination stays open

The updated skill removes a small-change bias. It distinguishes an idea’s natural ambition from the next affordable action.

Reports tell the truth

Work state, disposition, availability and measured value describe different things. A launch is not automatically a success.

What others have actually tried

Vendor accounts establish reported practice, not independent proof of ROI or transfer to our stack. Dates below identify the account, not a promise about every current account or plan.

Stripe
February 2026

Minions take a developer request to a CI-ready PR without intervening interaction, with human review. Stripe combines agent judgment with deterministic steps and caps CI repair rounds. Useful precedent for bounded delivery; it does not demonstrate autonomous product selection. Implementation detail.

Ramp
2026 case study

Sheets agents investigate monitor alerts and prepare fixes. Noise and duplicate alerts became a major problem; triage and a retained PR pointer helped later agents stand down. Detection, fresh ideas and a reason to act need different judgments.

Cursor
February 2026 research

A self-driving-codebase experiment exposed lock contention, a bias toward easy tasks, overloaded coordinators and premature completion. Its browser was a research project, not a production-quality delivery claim. We should borrow ownership and observability lessons rather than copy its scale.

Anthropic
November 2025

Compaction alone did not sustain reliable long work: agents lost unfinished state or declared success early. Durable progress, clear acceptance and usable handoffs helped. That supports our records and checkpoints; it does not justify prescribing a menu of ideas.

OpenAI & GitHub

OpenAI exposes a reusable harness for context, tools and configured approvals. GitHub separates agent reasoning from vetted writes and credentials. The portable lesson is to reuse capable runtimes while keeping permissions, admission and durable records explicit.

Cognition
May 2026

Async work increases the need for visible end-to-end proof. Reported failures included testing unrelated paths, getting stuck in setup and missing the core behaviour. Ground tests in the actual interface and intended journey; a clean review alone is insufficient.

StrongDM
February 2026

A more aggressive software-factory model removes human code writing/review and relies on scenarios and behavioural service twins. Its own account describes agents taking shortcuts to pass tests. This is a useful validation lesson, not authority to remove our existing gates or copy its spending model.

Sakana & Google DeepMind

The closest evidence for creative discovery: propose, debate, revise and empirically test hypotheses. Sakana’s October account explicitly acknowledges underdeveloped ideas and implementation errors, and uses grounded evaluation and shared findings. These systems still operate within research goals and evaluators; transfer to open-ended product value needs testing.

METR
February 2026 update

Its earlier trial found slower work with early-2025 tools; later measurements were confounded by selection and concurrent-agent timing. Neither that historical slowdown nor vendor PR counts establishes today’s ROI. Measure accepted outcomes and human effort here.

A simple next operating model

A nudge to explore→An idea—or none→A worthwhile next action↔Evidence & reflection

A conceptual loop, not a mandatory pipeline. Research, building, parking or stopping can be valid stopping points. The vocabulary remains provisional.

One shared project truth

Start with one logical dispatcher and common records. Codex or Claude can be an interface; multiple independent controllers should not maintain competing ownership or quota decisions.

Several agents can propose ideas. Shared admission claims the work, checks aggregate limits and distinguishes useful independent exploration from duplicate implementation. One writer owns each checkout.

Keep the trigger deliberately simple

Begin with manually requested exploration, then consider one approved native schedule once its prompt and cost are reliable. A timer requests admission; it does not compel a mission or a deliverable.

Codex local scheduling needs the computer/app available. Claude offers different session, desktop and cloud options. These are not interchangeable capability or quota guarantees. No Muse adapter is certified by this review.

A Git log is useful durable evidence, but a shared file is not a distributed lock. Start with serialized claims; add transactional leases only when real concurrency requires them. No AWS infrastructure is needed for the next supervised proof.

“Where is it?” and “Was it worthwhile?”

Work state	Exploring · Shaping · Making · Checking · Waiting · Parked · Closed
Disposition	Delivered · Useful learning · Not pursued · Abandoned · Superseded
Availability	Source · Local · Staging · Production · Withdrawn · Not applicable
Observed value	Unknown · Measuring · Positive · Mixed · Negative, with impact and evidence described in context

Proposed registry fields, not implemented live statuses. Keep an assessment date, evidence and a next observation where relevant. “Essential maintenance” can be valuable without exciting analytics; “breakthrough” needs demonstrated consequence.

Park at any stage. Keep the open question, artifacts, owner and revisit condition. A later authorized run can mull, research, seek approved advice, escalate, resume delivery, leave it parked or close it. Parking is not abandonment; ongoing work still has a visible budget, and resumption preserves the earlier evidence and limits.

Fun labels can add colour after the facts. “Lick of paint” describes a change, “bolt of lightning” an origin, and “long shot” a bet. They should remain optional descriptions, never categories the ideation agent must fill. “Curveball” or “backfire” reads more clearly than “IED” in a shared report.

Write the update people actually need

Products and About now make it clearer which page you’re on. A small usability fix, checked in source; user impact hasn’t been measured.

Lead with the change and its significance. Agent counts, test totals and timing support the account below the headline. A genuinely consequential result deserves a stronger update when the evidence warrants it.

Preserve the original closing record, then append later benefit assessments. If analytics are still being collected, name the measurement, owner and next review point; do not claim monitoring exists until it actually does.

Room to explore, with an owner-set budget

Ordinary work preserves the reserve. An explicit exploration burst near a reset could fund more breadth or depth. It must name the actual allowance, actors, duration and expiry; it does not automatically lift the current 80% stop.

A predicted reset is not an observed reset. Extra usage does not relax privacy, review or release gates, and no idea remains a legitimate result.

What to prove next

Carry the pilots’ discovery evidence forward. The next proof is low-babysitting initiation, safe shared admission and continuity across runs. Rehearse that path before enabling one explicitly approved schedule; verify empty starts, collisions and quota holds without inventing productive work. Measure total effort, rework and eventual usefulness—not just how much ships.

Keep one active implementation per project initially. Expand concurrency, schedules or release autonomy only in response to observed need and verified capability.

Who shaped this, and how it evolved

Initiator and contributor: Dean—creative intent, unbiased discovery, proportional reporting, outcome distinctions and owner-triggered budget bursts. Research and implementation: Codex. Source-review actors will be recorded only when actually observed; no independent research evaluation is claimed here.

The starting question asked what companies have achieved from ideation through deployment.
Comparing execution systems with scientific discovery clarified two different evidence questions.
Dean’s corrections kept our own demonstrated discovery in view and added outcome/value distinctions.
The skill was revised to preserve idea ambition, actual contributors and evolution; a registry/scheduler design remains a proposal.

Evidence, cost and limitations

Primary pages were checked on 29 October 2026. The comparison is a targeted review, not a systematic survey or an independent audit of vendor claims. Existing pilots are real local evidence with small sample size and incomplete total-cost/benefit coverage.

The post-admission native sample covers 09:01:03–09:16:21 UTC: 5,525,767 input tokens, including 5,446,784 cached; 78,983 uncached input and 19,858 output. It excludes initial admission, later reporting/PR/final output and reviewer usage. Counts are not a bill. Shared weekly usage was 28% at admission and 29% at this source checkpoint; that delta is not attributable task cost.

Provisional author assessments, 0–3: usefulness 2, correctness evidence 2, efficiency 1, observability 2, independent verification unknown, learning 2. No composite score. Source retrieval produced avoidable navigation output; targeted retrieval improved it. The findings support a next test, not a proven end state.

Origin/evolution · Score evidence · Measurement scope · Operational audit