@jholtdigital: A friend encouraged me recently if I want to know how well something works for my use case, I should try it for a month…
Summary
A user shares experience using FactoryAI to convert a design system from HTML/CSS to Flutter widgets with E2E testing. The tool employs an orchestrator, workers, and validators using multiple AI models to plan and execute long-horizon tasks over 79 hours, spawning over 229 agents.
View Cached Full Text
Cached at: 06/23/26, 09:54 PM
A friend encouraged me recently if I want to know how well something works for my use case, I should try it for a month. So I gave this a shot with another AI tool I’ve been wanting to explore.
I ran my first @FactoryAI Mission–a very cool idea from the Factory/Droid team. The structure they built for this kind of long horizon task – an orchestrator plans, asking clarifying questions, creates milestones, plans around your codebase constraints, and executes long horizon work. It’s a much more structured system than the /goal commands I’ve used with other apps.
Some stats from this run:
- Active time: 79h 2m
- Avg per milestone: 5.3h
- Avg per feature: 30.2m
- 229+ agents spawned–not counting subagents that they spawned (subagents for reviewing the code in areas like simplicity and architectural fit)
- 158 tasks
This specific task was taking a design system I had prototyped and fine-tuned in HTML/CSS and converting it to Flutter widgets with E2E testing in Patrol. It replaced the existing basic material design theme and widgets. The agent first had to capture a number of screenshots of the HTML prototype before planning the implementation. The initial plan that the orchestrator created was done by Fable 5–but post-plan creation the orchestrator was either Opus 4.8 or more often GPT 5.5 xhigh. Workers were Composer 2.5, some GLM 5.2, and some Kimi K2.6/K2.7. Validators were almost exclusively GPT 5.5
I’ve never seen a system like this. It’s very impressive. I need to spend the next couple of weeks doing a lot of refinements. The system is complete and working well but there are a number of UI mistakes that did not match the prototype correctly. I think though this would be the case no matter what. Agents can get you about 80% there and the last 20% of refinements is where a big chunk of the work is. Super impressed! Maybe I can do a longer write up if people are interested in more details.
Similar Articles
@rohanpaul_ai: Factory 2.0 is here. Connects AI agents to the whole software workflow: tickets, customer requests, code, tests, securi…
Factory 2.0 connects AI agents to the entire software workflow, managing a feedback loop from tickets to incidents, treating every signal as training data.
@PrajwalTomar_: Honest take: the part that got me was watching it work. It opened a browser, clicked through the site like a real analy…
The author built an AI employee using Hyperagent and open-source models like Kimi K2.6 to automate client research, generating pre-call briefings. Multiple models are available for swapping.
@rabois: Accelerate.
A tweet praising Factory's AI coding automation tool, highlighting its ability to handle low-leverage work and an auto-model router that automatically selects the best model among new releases.
@LinusEkenstam: One month with an AI employee in our workspace. Here is the honest ledger, including the one job I still refuse to hand…
A builder shares a one-month honest review of Viktor, an AI employee that saves their team ~12 hours weekly and caught a contract deadline, arguing the tool pays for itself.
@dabit3: Underrated feature of advanced cloud agents like @devinai: Infinitely parallelizable human-grade UI testing. They - Sta…
This tweet highlights that advanced cloud agents like Devinai can perform infinitely parallelizable human-grade UI testing by setting up environments, inspecting code, planning e2e tests, handling authentication, using computer vision and mouse-and-keyboard control, and returning annotated video proof, with reusable skills for faster future verification.