GPT-6 Astra's first Choice batch: three builds and a coordinating agent that caught an input bug

Three agents, three self-chosen builds

A dev.to post by billiem documents the first "Choice batch" from GPT-6 Astra: three agents, each given latitude over its subject, delivered a string puzzle, a supply-chain teaching model and a tool for imposing a foldable booklet, published on 4 September. The notable part is less the artefacts than the workflow around them — a coordinating agent reviewed each prototype before polishing began and, in one case, spotted an input defect in the puzzle and returned the build to its builder while the game was still unfinished.

According to the post, the author requested three independent experiments with room to choose the subject. Each Astra agent ran at maximum reasoning under a broad remit — something to play, explore or use. The standing brief required prototype review before polish, and the batch's curation record logs the corrections each reviewer asked for. The author is explicit about the limits of the exercise: these were separate projects with fixes applied, not controlled comparisons against another model. The dev.to piece is itself an adaptation, made with AI assistance, of a longer article on billiem.uk.

Loose End and an input bug caught mid-build

The puzzle, Loose End, asks the player to steer the brass tip of a finite string across a board holding three pegs and four bells. Because the string catches on the pegs, the route taken to one bell determines how much length remains for the next, and ringing all four bells is not the win condition on its own — the end has to travel back home.

The first prototype's arrow-key controls ignored short taps. The coordinating agent found the problem and sent the build back before it was complete; the fix made brief keypresses register. Later checks went beyond a single win: they exercised two distinct winning strategies and retraced a long sequence of moves back to its starting point. The post is careful about scope, noting that these checks cover the game's route and unwinding rules only — the underlying planar taut-string model does not simulate friction or stretch, and the game consists of one authored course.

The Order Echo: two ordering policies on one scale

The second build models a shop, wholesaler, depot and factory reacting to the same demand under different ordering rules. The reviewer's request concerned visibility: the traces should sit on a shared scale, and the contents of each order should be inspectable. The finished interface lets you pick a week and a stage to examine its arithmetic.

Under the starting policy, one busy shopping week — twelve crates instead of four — drives the chain's factory orders to a peak of 76.8 crates in week twelve. When the policy instead counts outstanding orders, including supplier backlog, the shop initially orders more, 14.5 rather than 12.5 crates, because the higher demand forecast raises its target and adds a correction — yet the factory peak falls to 38.1 crates in week nine. The starting policy's traces remain visible as dotted lines beside the changed run, so result and timing can be compared directly. The post stresses that all quantities belong to the model's chosen rules — a 48-week window, continuous crate amounts and unlimited production upstream of the factory — making it a teaching model rather than a business forecast.

Pocket Press: a foldable booklet, untested on paper

The third tool takes a cover, manuscript and back cover and arranges an eight-page booklet on a single A4 or US Letter sheet, keeping a reading view alongside the print layout. It can save printable output, an SVG sheet and editable source. The reviewer asked for discoverable save actions, folding instructions and a way to reopen saved source; recorded browser checks downloaded the actual files, read back their geometry and text, and restored the manuscript from its saved source. Overflow blocks printable exports while leaving source saving available.

The imposition is deliberately unfamiliar — pages 5, 4, 3 and 2 upside down above 6, 7, 8 and 1 — meant to become a book after cutting a central slit and folding. Capturing images for the article surfaced a separate flaw: at a 1,280-pixel browser width on 9 September, the print sheet extended past its preview area and was cropped at the top and bottom, while the 900-pixel view fits. The paper result remains untried — every check was run by agents or automation, with no printer or fold involved, and the tool ships a numbered test sheet for the part of the experiment that needs a physical page.

Why it matters

This is an early, well-documented look at multi-agent coordination in self-directed builds: a reviewer role embedded in the brief caught an input bug before completion, requested interface changes that shaped each tool, and left a curation record of every correction. It is a demonstration of process rather than a benchmark — a single batch with no baseline model — and validation stayed inside the browser, with the physical booklet untested. Still, for anyone evaluating agentic coding workflows, the interesting signal is how review-before-polish changes what agents produce, and how openly the limits were reported alongside the results.