# Useful work. Recovered. Reused.

## AGI Alpha — advanced computer-use study

**14 October 2026 · Public report edition · Executed engineering study with a simulated nontechnical owner**

[Read the visual report](https://montrealai.github.io/AGI-Alpha-Agent-v0/reports/computer-use-study/) · [Download the public report pack](https://montrealai.github.io/AGI-Alpha-Agent-v0/reports/computer-use-study/report-pack.zip) · [Inspect the recorded results](results.json)

> **The practical finding:** AI-assisted work produced two repaired websites. Saved work was recovered in the tested scenarios, a retained repair prevented the same failure on a new project, and saved preferences reduced repeated setup. The owner was simulated; the software and browser checks were executed.

## The results at a glance

| Question | Recorded result | What it means |
| --- | --- | --- |
| Can it produce useful work? | Task A: **31/31** checks; Task B: **35/35**. | Two repaired bilingual workshop websites passed their declared mechanical checks. |
| Can saved work be recovered? | **19/19** recovery and refusal checks. | Repeatable restoration and defined refusals in a controlled environment; not a production reliability rate. |
| Does a retained repair help a new task? | Task B: **35/35** with the repair, **29/35** with that one repair removed. | One recovery-export defect was prevented. The six failed checks share that cause. |
| Can repeated setup be reduced? | **14 → 11 controls**, or **21.4% fewer scripted operations**. | Less configuration in the tested interface, with the same final quality, current authorization and review checks. |

These results come from the attached study's records, not from a new evaluation of the current repository. The primary evidence is reproduced in [results.json](results.json), including the ten owner replays, nineteen recovery checks, original protocol and source hashes.

## 1. A nontechnical owner's request

The fictional owner operates public workshops and does not write code. The request is deliberately practical:

> Repair my workshop website. Keep the offers, fix the estimates, make French and mobile use work, give me files I can keep, and let me recover safely. Do not publish, book, send or pay anything.

Task A, **Atelier North**, has three offers and a 10% discount beginning at ten participants. Task B, **Studio Horizon**, has four different offers, a free offer, a sold-out offer, a six-seat offer, longer French titles and a 15% discount beginning at eight participants.

The work products are usable HTML/CSS/JavaScript applications, reviewed quote files, recovery examples and retained implementation methods. This is plausible business work, but real customer acceptance, revenue and willingness to pay were not measured.

## 2. What the AI actually did

The conversation assistant authored source code, directed browser interactions, inspected screenshots and corrected observed failures. The workflow combined reasoning, code generation, structured file operations, keyboard/mouse interaction and deterministic verification. It did not force every operation through screen coordinates when a direct tool was more appropriate.

Browser actions used Chromium through Playwright. Exact arithmetic was checked separately in Python using Decimal and half-up rounding. That oracle is a separate implementation, **not an independent organization or reviewer**.

No separate live model API benchmark, metered computer-use provider session, independently operated reviewer model or weight-training run was performed. Reproduction scripts replay mechanical tests; they do not recreate the original model's reasoning or independently generate another repair.

## 3. First task: useful work, visible mistakes, a tested repair

The starting Task A website was an intentionally defective fixture. It had mobile overflow, overlapping header text, no French experience or recovery, missing review gates and incorrect boundary behavior. It passed **8/31** checks. These were seeded fixture defects, **not defects discovered in AGI Alpha's source**.

A first coordinate-selected language click missed its target. Inspecting the screenshot and correcting the click allowed the next interaction to proceed. A subsequent quote calculation succeeded, but recovery export failed because Web Crypto was unavailable in the memory-only browser context.

The repair added a portable SHA-256 implementation for that context. All **105** recorded UTF-8 vectors matched Python hashlib, including empty, non-ASCII and longer inputs. This verifies agreement on those vectors; it is not a cryptographic audit.

The final Task A website passed **31/31** checks. A representative result was 10 × CAD 49.50 = CAD 495.00, less CAD 49.50, yielding **CAD 445.50**. No booking, payment, tax calculation or message occurred.

Both observed mistakes remained in the study record. An attractive result should not erase the failures that explain how it became reliable within its tested scope.

## 4. Next task: a different project, a retained method

Task B changed the brand, catalogue, identifiers, prices, capacities, French content and discount policy. Those facts and permissions were not inherited from Task A.

| Task B comparison | Passed | Interpretation |
| --- | --- | --- |
| Deliberately defective starting fixture | 9/35 | The supplied starting condition. |
| Repaired application using the retained method | 35/35 | All declared B checks passed. |
| Identical repaired application without the portability repair | 29/35 | One root cause prevented six recovery-dependent checks from succeeding. |

The last comparison is a **component-removal test**, also called an ablation. It isolates one real repair. It is not a weaker-model baseline, a six-defect improvement or proof of general recursive self-improvement.

Fresh cases included an eight-place total of **CAD 433.50** and a rounding case in which CAD 95.625 was rounded to CAD 95.63, producing **CAD 541.87**. The free offer stayed at zero and the sold-out offer could not produce an estimate.

The retained profile did **not** improve mathematical correctness over the same fully repaired application without saved preferences. Both had 35/35 final acceptance. The profile's benefit was less repeated configuration.

## 5. Recovery: retain work, require new review

The study tested two distinct mechanisms. A project checkpoint retained exact source and brief bytes and checked a separately retained latest head. A browser draft restored selected inputs, checked catalogue binding and required a fresh calculation and review.

The **19/19** checks comprise one interrupted worker, five fresh-process artifact restorations, five fresh-browser draft restorations, and eight refusal/authority checks. The refusal cases cover tampering, an older checkpoint compared with a newer retained head, a wrong task, an unknown-effect marker, a parent-directory path, overwriting existing work, inherited authority and missing unsaved work.

All five artifact restorations were byte-identical. All five fresh-browser restorations produced the same reviewed quote bytes after new review. The worker was deliberately terminated after checkpoint creation; this was **not a power-loss experiment**.

### What recovery does not promise

Unsaved work can still be lost. A checksum is not identity, encryption or proof that a claim is true. Freshness requires the latest trusted head outside the recovery package; the example head is supplied alongside the example only for reproducibility. Browser drafts have no independent freshness anchor.

The unknown-effect marker was synthetic. No live external transaction was reconciled. An unresolved effect must be inspected before a new action is authorized; it must not be blindly replayed. Hardware failure, malicious hosts and cross-operating-system disaster recovery were not qualified.

## 6. Less repeated setup — not a human productivity claim

The owner-interface comparison used the **same Task B, same finished application and same evidence**. Five familiar/no-profile replays and five retained-profile replays were executed in alternating order. These were ten automated runs, not ten people or ten independently generated agent solutions.

| Interface obligation | No saved profile | Retained profile |
| --- | --- | --- |
| Total recorded controls | 14 | 11 |
| Preference entry | Four separate edits | One profile confirmation; zero edits |
| New current-task scope approval | 1 | 1 |
| Artifact review checks | 3 | 3 |
| Final Task B correctness | 35/35 | 35/35 |

The reduction is `(14 - 11) / 14 = 21.4286%`, rounded to **21.4%**. No permission or review gate was removed to produce the saving. The same 180-character brief was entered in both conditions; no typing reduction is claimed.

This advantage depends partly on the interface. A well-designed batched preference entry could reduce or eliminate it. Attention, confidence, comprehension and actual human time were not measured.

### Modeled time is separate

The original editable model assigns 120 seconds to brief/scope, 180 to review and 30 to keeping files in both conditions. It assumes 100 seconds for four preference choices versus 35 seconds for confirming saved preferences: **430 versus 365 seconds**, an assumed 65-second saving. These are scenario inputs, not stopwatch observations. Under the study's stated ranges, the assumed saving spans **0–160 seconds**. No amortized return or provider-cost advantage was established.

## 7. How much confidence should a reader place in it?

**Executed:** generated application behavior, browser controls, quote downloads, component-removal checks, recovery checks and file-hash comparisons.

**Simulated:** the owner, workshop businesses, input catalogues and unknown external-effect marker.

**Modeled:** human-time assumptions and any interpretation of business value.

**Not established:** real-user savings, broad autonomous competence, strongest-agent superiority, independent scientific replication, comprehensive accessibility, production security or a commissioned live marketplace.

Both tasks were small and constructed. The same assistant knew the briefs, authored the defective fixtures, wrote the repairs and designed the checks. The primary protocol preceded acceptance runs, but the detailed harness followed the initial visual/export probe. This was neither blinded nor externally preregistered. B's requirements were not hidden even though the retained template was frozen before the repaired B build.

## 8. Environment and repository boundary

The recorded environment used Python 3.13.5, Playwright 1.57.0 and Chromium 144.0.7559.96. Fresh browser contexts blocked HTTP/HTTPS during fixture tests. The environment denied file navigation, so the exact HTML bytes were supplied through `page.set_content`. First-time double-click opening from disk was not qualified in that study. Chromium required `--no-sandbox` inside the surrounding container; that is not a production-isolation recommendation.

The source-scope review referenced repository commit **8c027575c850bf5a872e6b11f81933b5e77b6be2** and observed release **v1.29.0** at that time. **The study did not execute the repository or change it.** Publication of this report does not turn those results into validation of a newer commit.

The pinned [Computer Use guide](https://github.com/MontrealAI/AGI-Alpha-Agent-v0/blob/8c027575c850bf5a872e6b11f81933b5e77b6be2/docs/agent/SUCCESSOR_COMPUTER_USE.md) describes a narrower isolated-browser adapter. The hybrid code-and-browser workflow studied here is an integration reference, not a claim that the pinned adapter already provided arbitrary code or desktop execution.

## 9. The useful next step

Add a separately authorized code-and-browser executor without widening existing saved permissions. Preserve artifacts and action intent durably, require new review after restoration, retain approved methods rather than authority, and keep all existing workflows and safeguards.

Then conduct a real-user pilot: counterbalance task order, use a capable same-model/no-memory comparison, and have a separate reviewer grade exact artifacts. Measure active attention, accepted quality, help requests, incorrect approvals, recovery, full costs and rework. Include first-task setup cost and negative transfer.

The original proposal suggested twelve participants, at least 90% unaided completion, no critical wrong approvals and at least 20% median reduction in matched-task active effort at equal or better quality. **These are proposed pilot targets, not achieved results or validated universal thresholds.**

## 10. Publication, downloads and provenance

This is a reviewed, accessible **public report edition**, not a new experiment or software release. The compact ZIP includes this Markdown report, the English/French visual reader, styles, reader controls, curated recorded results and publication notes. It does **not** contain the original archive's PDF/Word/Excel files, screenshots, complete applications or full reproduction harness. Do not describe the compact pack as the complete original study.

To retain a readable copy, download the public pack and open `START_HERE.html`. To produce a PDF, use **Print / Save PDF** in the reader. No account, wallet, provider key, analytics, remote font or model is required to read it. Ordinary hosting requests still reach the hosting service.

The original archive was supplied as `AGI_Alpha_Advanced_Computer_Use_Study_2026-10-14.zip` (**3,632,846 bytes**). Its SHA-256 is:

`0e9d8a7f3d163f2461af99261d11f464ead71411ae9803074171ac934a123035`

All **97** manifest-listed files in that supplied archive were checked before this publication. Its original report Markdown SHA-256 is `4abf31e92d89aea25d3ec1e6cbc834817a5f80a34236177e21aacb58a2f6a82b`. The web edition has different bytes and is deliberately identified separately.

[Results and source fingerprints](results.json) · [Publication notes](PUBLICATION.md) · [Back to the project](https://github.com/MontrealAI/AGI-Alpha-Agent-v0)

---

**In one sentence:** useful work can leave behind more than an answer—a checked artifact, a recovery path and a method worth testing again.