Agent Canvases Make Work Visible; playagit’s Publication Counts Still Need Explaining

Evidence note: This article draws on vendor descriptions, vendor-hosted customer accounts, and playagit’s own recorded pipeline measurements. External capability and performance claims remain independently uncorroborated here. Pipeline figures describe playagit’s records, not an independent audit; calculations are identified as derived analysis.

Visible plans, review controls, execution records, and model costs answer different questions about an agent workflow. None alone establishes how much finished work reached its destination.

That distinction matters in playagit’s own records: across 172 runs from August 28 through October 13, 2026, the pipeline reported 1,360 published posts, including 376 new and 379 updated posts. Those categories do not reconcile with the reported total. The immediate finding is an accounting gap, not evidence that the unexplained publications succeeded, failed, or belonged to any particular category.

Azure Canvases and Ubik: shared plans, document comparison, and approval

Microsoft’s September 29, 2026 description presents Azure Canvases as shared interfaces where users can inspect progress, make changes, and approve next steps alongside Copilot. However, the broader claim that plans, code, previews, and deployment feedback form a complete collaborative development workspace remains uncorroborated in the supplied evidence. Microsoft’s Azure Canvases description

For Ubik, the supplied evidence ledger lists claims that Compare Canvas supports side-by-side text comparison and selective changes, and that its context-aware workflows incorporate human approval. Both remain open: the linked homepage currently displays a winding-down notice rather than documentation establishing those capabilities. Ubik’s homepage

The practical evaluation question is whether approval binds to the exact document or change that will be applied. A visible proposal is useful, but assessing a review control also requires knowing what action it authorizes and whether subsequent changes invalidate that approval.

Docker Sandboxes and Jump Trading: execution evidence and human review

Docker’s September 2, 2026 article describes an evaluation kit that delegates commands to Docker Sandboxes and records execution details, including output, exit status, and elapsed time. Its claimed contribution to reproducibility remains uncorroborated here; the article also explicitly distinguishes command execution and evidence capture from running models or automatically judging their results. Docker’s evaluation workflow account

OpenAI’s October 6, 2026 customer account describes Jump Trading combining multiple data sources in extended research workflows, with human review and acceptance of outputs. That remains a reported operating practice rather than independently established evidence of the workflow’s effectiveness. OpenAI’s Jump Trading account

These examples suggest separate evaluation criteria: execution records should establish what ran, while review records should establish what a person accepted. Neither should substitute for evidence that the result met its intended purpose.

Selective compute: planning location, estimated savings, and browser tests

Google’s September 23, 2026 Antigravity SDK article describes a hybrid demonstration in which a cloud model plans work and local models perform auditing and patching. This remains a vendor-described capability, with its effectiveness outside the demonstration unresolved in the supplied evidence. Google’s Antigravity SDK announcement

OpenAI’s October 8, 2026 LegalOn account reports a 65% reduction in estimated daily Codex costs, alongside maintained development speed, after changes to model selection and budget management. Neither the savings nor the maintained-speed claim is independently corroborated here; estimated daily costs should not be presented as audited expenditure. OpenAI’s LegalOn account

OpenAI’s October 9, 2026 Asana account reports 76-fold lower estimated model costs and fivefold faster browser runs in tests. Those results remain independently uncorroborated. The account distinguishes GPT-6 Astra in Codex, which investigated and tested improvements, from GPT-6.1 Sol, which powered the optimized browser workflow. Treating the result simply as a faster model would obscure that distinction. OpenAI’s Asana account

For developers, these are different propositions to evaluate: where work executes, how tasks are assigned, and how workflow changes affect cost and completion time. A useful comparison should hold the task and acceptance criteria stable while recording retries, review effort, and the cost of accepted results.

Pipeline records, August 28–October 13, 2026: discovery and production selection

PlayAgit’s own records for 172 runs from August 28 through October 13, 2026, as of October 13, report 34,443 discovered topic candidates across 4 categories and 223 distinct source hosts.

For that same 172-run period, playagit’s records report 40,792 scored topics, with 15,859 held before production, 17,599 rejected, and 7,334 sent to production.

The recorded scoring dispositions reconcile arithmetically:

15,859 held + 17,599 rejected + 7,334 sent to production
= 40,792 scored topics

This is a derived consistency check on playagit’s supplied totals. It does not establish why the scored-topic count exceeds the discovered-candidate count. The supplied summary does not define a mapping between those counters, so a discovery-to-production conversion rate would require additional assumptions.

Research coverage across the recorded runs

PlayAgit’s own records for 172 runs from August 28 through October 13, 2026, as of October 13, report 5,501 research dossiers containing 30,901 claims, of which 1,825 were independently cross-checked and 7,006 were supported by primary sources.

These labels describe different evidence relationships. Primary-source support establishes what an originating organization says; independent cross-checking describes an additional verification process. The aggregate records do not establish whether the categories overlap.

Consequently, subtracting both categories from the claim total would not establish a reliable count of unsupported claims. Nor does the summary show which research claims ultimately appeared in published articles. Coverage across dossiers and coverage within published work remain separate questions.

The draft gate through October 13, 2026

PlayAgit’s own records for 172 runs from August 28 through October 13, 2026, as of October 13, report 1,576 draft reviews: 1,307 publication decisions, 170 revision decisions, and 99 rejections.

The recorded decisions reconcile:

1,307 publication decisions + 170 revision decisions + 99 rejections
= 1,576 draft reviews

This derived check establishes agreement within the review tally. It does not establish that every review concerns a unique draft, or that every publication decision resulted in a live post. Those conclusions would require records connecting drafts, decisions, and publication outcomes.

The publication breakdown accounts for only part of the reported total

PlayAgit’s own records for 172 runs from August 28 through October 13, 2026, as of October 13, report 1,360 published posts, with 376 classified as new and 379 classified as updated.

The supplied categories produce this derived subtotal:

376 new + 379 updated = 755 classified publications

That is the central limitation of the breakdown: its named categories do not account for the full reported publication count. The records supplied here do not establish whether the total and categories use the same counting unit.

Related reading includes playagit’s earlier analysis of publication totals that exceed their accompanying breakdown. That article provides background, not evidence reconciling this reporting period.

What the arithmetic leaves unresolved

Using playagit’s supplied publication figures for the 172 runs from August 28 through October 13, 2026, the derived remainder is:

1,360 reported publications − (376 new + 379 updated) = 605

The remainder of 605 is an arithmetic result, not an observed publication category. It cannot be labeled as duplicate posts, retries, manual publications, failed writes, or any other activity from the supplied evidence.

The same reporting period also has different recorded totals for publication decisions and published posts. No supplied reconciliation connects those counters. Their difference alone does not demonstrate that a review gate was bypassed.

Resolving this requires definitions of each counter, a consistent reporting cutoff, and records connecting each counted publication to its content identity, review decision, and final outcome. Whether the underlying records already support that reconciliation remains open.

Developer takeaway: evaluate controls, execution, and costs separately

A practical assessment should ask distinct questions:

• Review controls: What exact artifact does approval cover, and which action does it authorize?
• Execution evidence: Can the record identify the environment, inputs, command, output, and completion status?
• Model costs: Does the comparison measure estimated or billed costs, and does it include retries and human review?
• Publication accounting: Does each counter represent a decision, an operation, or a unique live article?

For publication systems, the immediate follow-up is a reconciliation using the same time window and explicit counter definitions. Until that exists, a dashboard can accurately display recorded totals while leaving their relationship unexplained.

How we apply this in Vision

The evidence-backed application to Vision is to keep selection, review, and publication measures distinct. PlayAgit’s own records for 172 runs from August 28 through October 13, 2026, as of October 13, record 7,334 topics sent to production; that figure describes a production-selection outcome.

PlayAgit’s own records for 172 runs from August 28 through October 13, 2026, as of October 13, record 1,576 draft reviews and 1,307 publication decisions; those figures describe the review stage.

PlayAgit’s own records for 172 runs from August 28 through October 13, 2026, as of October 13, separately report 1,360 published posts, 376 new posts, and 379 updated posts; those figures leave the publication breakdown unresolved.

For Vision, these records support a reporting requirement: define each measure and make its relationship to neighboring stages inspectable. They do not establish that such reconciliation has already been implemented.

Related on PlayAgit

• Copilot Review Access and the 565 Publications Playagit's Breakdown Does Not Explain
• 1,423 Draft Reviews and 1,231 Publications: What playagit’s Pipeline Counts Explain
• @huggingface/kernels: 200+ WebGPU Kernels for Running AI Locally
• Zero-Trust Architecture for Autonomous AI Agents That Touch Production