Coding agents write code fast, then hand you more to supervise: diffs to review, previews to test, follow-ups to write. Coroid is an AI software factory that takes that work off your plate and hands your team a pull request that has already been built, tested in a running browser and reviewed.
Create a 3D scene of a city block. Emphasis on detail is very important. The scene must have a timeline slider in the top, with the following options: 1945, 1965, 1985, 2005 and 2025.
ArchitectDone
DeveloperDone
QADone
ReviewerDoneQA sent the first build back with two failed criteria. The developer fixed both and QA verified them again before the reviewer saw the work. No one prompted them in between.
One task nobody had to drive
6 agent runs, 2 delivery cycles and all 11 acceptance criteria verified in under two hours, for 34.7m tokens and US$1.00 in model spend.
Recreated from a real task in Coroid. The screenshots are the QA agent’s own captures. The prompt comes from Bijan Bowen, who uses it to test AI models.
Read the Chrono City runA session, local or in the cloud, needs a person to run it, and the more you hand it, the more it hands back. You stay the bottleneck. Of the four jobs below, Coroid takes three and leaves you the one that should stay yours: the final call.
The job
In an agent session, local or cloud
In Coroid
Coroid is a software factory, not a better coding agent. It runs the line around the code, so your team’s attention goes to what gets built and what ships.
In an agent session, you are the QA person for your own automation. You open the browser, click every year, spot what is broken and write the follow-up prompt, round after round. Here the QA agent did that work and sent the build back before anyone was asked to review it.
Acceptance criterion 3
“Selecting 1985 renders no city: the viewport is a flat purple film-grain field while the HUD reports 1985. Reproduced three times.”
1985 renders its full neon skyline, whether it is picked on the slider or with the 3 key.
Acceptance criterion 6
“Clicking a building flies the camera in and opens the card, but the card contains no description.”
Every building, vehicle and pedestrian card carries period copy, and a new test fails if one is ever empty again.
The architect estimated 12m tokens. The run used 34.7m: the developer checked its own renders, QA exercised every era in a real browser, and one rework round fixed what QA found. With 97% of those tokens served from cache, the whole task cost US$1.00.
1945
1965
1985
2005
2025
It takes longer than a demo. That is the point: you get a pull request that has already been tested, not one you still have to test.
The same software factory handles a one-line fix and a multi-week plan. You stay where it matters: deciding what gets built, and whether it ships.
Screens are recreated from the portal. The plan’s totals are from a real run; the other example work is illustrative.
Write what you need in plain language, attach screenshots or notes, and pick a project. Coroid turns it into a specification with user stories and acceptance criteria before any code is written.
Finance teams keep asking for their invoices in a spreadsheet. Add an export button to the invoice list that downloads the current filter as CSV.
Done when: the file opens in Excel and Numbers, amounts use the account currency, and exports over 10,000 rows don’t time out.
Larger requests are broken into phases and tasks, with dependencies, estimates and a viability check. This 3D city scene was built as five phases and fourteen tasks in just over six hours.
Create a 3D scene of a city block. Emphasis on detail is very important.
The scene must have a timeline slider at the top with the options 1945, 1965, 1985, 2005 and 2025.
Selecting a year transforms the block in front of your eyes: buildings, vehicles, storefronts, advertising and the pedestrians’ clothes.
Ready work is picked up in priority order as agent slots free up. You see what is running, what is next, and anything waiting for you.
The Coordinator reads the repository and the specification and writes the plan, including the acceptance criteria that decide success.
It hands the work to one specialist at a time, with clear instructions. The developer builds, QA verifies the result, and the reviewer reads the diff.
When QA fails the work or the reviewer asks for changes, the Coordinator sends the fixes back to the developer, and QA checks the new version.
The task can finish only when QA has passed the exact version that will ship and, where review applies, the reviewer has approved it. The Coordinator cannot overrule either verdict.
Requests arrive from people, your product, and your integrations.
Coroid sanitises, deduplicates, verifies, and routes the request.
Qualified work enters the governed production line.
One Quality Graph gates the pull request, release, and canary.
You review evidence, not prompts. This is attached to every piece of work Coroid hands you.
How specifications guide the workEach task runs in its own cloud workspace with access scoped to one project. Agents cannot reach your network or production.
Pick an AI profile per task or per agent, bring your own provider keys, or keep inference in the EU.
Agents do the work; people approve it. Your quality policies decide what must pass before anything merges.
Keep them. They make each engineer faster at writing a change. Coroid covers what they leave to a person: choosing the next piece of work, testing it in a running build, reviewing it against your policies and keeping the record. Your engineers keep their tools and review what Coroid hands them.
You could. You would need specifications with acceptance criteria, a queue with capacity limits, browser QA that sends work back, review policies, evidence bound to each commit, spend per task and an audit trail, plus someone to keep it all working as models change. That is a software factory, and it is what Coroid runs. Most teams would rather spend that effort on their own product.
No. It takes on well-specified work so your developers can spend their time on direction, architecture and review. Every change still needs your team’s approval before it merges.
You pay for agent capacity and for the model tokens the work uses, at provider rates with no markup. The Chrono City task shown here used 34.7 million tokens and US$1.00 in model spend. Larger work costs more, and every task shows its own spend.
Each task is verified against its acceptance criteria in a running build, reviewed by a separate agent, and held to the quality policies you set. Work that fails goes back with the reason, and you only review what passed.
No. Connect a GitHub or GitLab repository in the browser. Agents work in Coroid’s cloud workspaces, so there is no IDE extension or local setup.
Into an isolated workspace for the length of the task, with access scoped to that project. An EU hosting profile and EU-only models are available, and on-premises deployment is in private preview.
Coroid Expo
Visit the Coroid Expo, a 3D trade fair where every feature has its own booth, and talk to the agents that run it.
Pick something your team keeps postponing, hand it over and walk away. Come back to a pull request that has already been tested.
