Agent Skill ยท public beta
Canvas apps, Dataverse solutions, Power Automate flows and Power Pages sites with the definition in git, a portable artifact built from it, a deliberate deployment, and every change proved by performing the task in the product, driven by Playwright.
What still stands between this release and the 1.0 criteria, in the order it is being worked. The same table is in the README and the roadmap; CI fails when they differ.
| # | Item | Why it matters | Status |
|---|---|---|---|
| 1 | Power Pages: a site walk driver with sign-in, refusal and /_api proofs, and a reference covering row visibility, "my records", Private sign-in, deploy and teardown | The thinnest area: written from one real site, and nothing could walk a site or prove a refusal on one | First measured build (0.21): 30 of 30 blind checks, design 23 of 25, 85 minutes, no person needed. 0.22 adds Dataverse confirmation, captures and freshness waits to the walker, and the build's lessons (go-live decisions, identity for people who are not Dataverse users, the 15-minute cache window, feedback patterns). Second measured build (0.22): 30 of 30, design 23 of 25, 96 minutes (30 lost waiting on the site's sign-in consent), no person needed; it used the bundled walker for all 65 checks. 0.23 accepts the site's own consent when nobody is present, adds spill and covered-field checks to the walker, and stops asking a build with no seed to re-seed. Third measured build (0.23.1): 30 of 30, design 23 of 25, 90 minutes, no person needed. 0.24 turns what blind reviews kept finding into checks (current-page marker, focus contrast, double writes, unfilled columns, unguarded Global read), holds a site's hand-back for a critique of 30/40 and a five-topic security review, settles identity first with the administrator's steps, and ships in one call. 0.25 scopes every rule to the surface it serves (rules-and-scope.md): the 30/40 floor is for sites only, and the flow lint gates every import; 0.25.1 reads server logic in the permissions audit. 0.26 adds what a per-person visibility build taught: several relationships deciding visibility, server logic's read cache, and the site address and rename procedures. Microsoft's plugin arm is next |
| 2 | Speed and cost against Microsoft's plugins | The last four canvas builds beat them on quality but not on time or cost | 0.21: data-source check on create, a design-helper time box, no single-issue publishes after the fix batch; to be measured |
| 3 | A self-test in CI for every script, and a floor that refuses to pass when nothing was examined | A 1.0 criterion; four scripts had none | Done in 0.21. 0.27 extends the floor: a text check that measured none of its bound text exits 2, and one that measured under half warns |
| 4 | Teardown order for Fabric: Dataflows before their lakehouse | Dataflows can refuse deletion once their destination is gone | Done in 0.21 |
| 5 | Stabilise (0.9): one interface table, the compatibility file, upgrade notes | Users need options, exit codes and finding codes that stop moving | Interface table and compatibility file drafted in 0.21; freeze at 0.9 |
| 6 | Re-run the graded evaluation on the current version: ten or more task types (adding Power Pages, model-driven and SharePoint), three runs each, the held-out triggering set | A 1.0 criterion; the last graded run was 0.7.1 (217/222 with the skill, 136/222 without) and much has changed since | Open |
| 7 | Publish the 0.7.x results and the measured builds on the evaluation page | The page still shows the 0.5.1 run | Open |
| 8 | A model-driven app with a responsive canvas companion, built and verified in a tenant | Clears most of the snippets still marked "confirm in your tenant" | Open |
| 9 | SharePoint as a data source: delegation, the 5,000-item threshold, staying on a standard licence | Common in real tenants; written down nowhere yet | Open |
| 10 | Approvals in Teams and Outlook, and the one-Approvals-connection-per-person limit | Every approval flow meets it | Open |
| 11 | Two independent real builds in different tenants | A 1.0 criterion; every build so far ran in one tenant | Open: needs a second tenant |
| 12 | Driver commands add-data and refresh-source for an existing app | Both were done by hand-written browser helpers in two projects; the recipe is in browser-verification.md section 12 | Open: port and prove against a live Studio session before release |
| 13 | SKILL.md against Anthropic's current skill guidance: body under 5,000 tokens with the must-hold rules first (Claude Code keeps only the first 5,000 tokens of a skill after compaction), and a measured check that the newest models do not need the prescriptive passages | The body is about 10,500 tokens; after compaction everything from the tool table down is lost | Done in 0.31: about 4,100 tokens, 141 of 148 graded checks as before, 9% cheaper; CI fails a body over 20,000 characters |
| 14 | Driver command insert-gallery --screen S --name N for the measured new-gallery workaround (Insert > Blank vertical gallery, rename, save) | Studio tears down on a push that adds a gallery to an existing screen; the workaround is measured twice but its Studio steps are still hand-driven | Open: port and prove against a live Studio session before release |
| 15 | A pre-import check that refuses a package whose flow StateCode differs from the repo | A client-side import timeout left active flows in Draft, and a package rebuilt from the live export would have carried the drift into the next import; only the post-import state diff exists today | Open |