AI · 3 min read
The secret we used to make our workflows 85x faster: subagents
Not a smarter model. Not a better prompt. Getting the AI to stop waiting its turn.
In short A workflow that took 3 days ran in 40 mins to an hour as a sequential AI workflow (24x faster), and is now 85x faster than the original. Two changes made the difference: the AI starts reading the floor plan the moment it is uploaded, while the brief is still being written, and a lead agent splits one job into many, handing each piece to its own subagent so they all run at once. Four renders: 321 seconds one after another, 82 seconds side by side.
- Published
Key takeaways
- The same workflow went from 3 days, to 40 mins to an hour as a sequential AI workflow (24x faster), to 85x faster than the original with subagents running in parallel.
- Not a smarter model or a better prompt: the gain came from getting the AI to stop waiting its turn.
- Start early: the AI reads the floor plan the moment it is uploaded, while the designer is still writing the brief.
- Split one job into many: a lead agent hands each render to its own subagent. Four renders took 82 seconds instead of 321.
- The steps keep their order. Subagents only work in parallel inside a step, and they all read from the same plan.
How much faster did the workflow get?
- Original timeline: 3 days
- AI sequential workflow: 40 mins to an hour (24x faster)
- AI parallel workflow with subagents (85x faster than original)
👀 We quickened a workflow that was already 24x faster to something that is now 85x faster.
Not a smarter model. Not a better prompt. Getting the AI to stop waiting its turn.
The whole workflow
- Original timeline3 days
- AI sequential workflow40 mins to an hour · 24x faster
- AI parallel workflow with subagentsabout 17 mins · 85x faster
Four renders
- One after another321 seconds
- Side by side, one subagent each82 seconds
Why do most AI workflows run like a waterfall?
🌊 Most AI tools / workflows work like a waterfall. You fill in a form, press submit, and only then does the AI start. It reads, then plans, then builds, then produces the output, each step waiting for the one before it.
That's how one person works alone. Software doesn't have to. 🤖
What happened when we built IDstudio?
🏠 When I was building IDstudio, which turns an interior designer's floor plan and brief into a furnished layout, a 3D model and realistic renders, we initially had to test it sequentially, debug, and run it again. Run as a waterfall, it took about 37 mins to an hour.
🚀 Now, with its functions running concurrently, it takes about 17.
Two changes made the difference. 👇
How can the AI start before the brief is done?
1️⃣ Start reading as the brief is being inputted, not only at the end.
The designer uploads the floor plan 📐, then spends a few minutes on the brief ✍️. Reading the plan doesn't need the brief: the walls, doors and dimensions are all on the drawing. So the AI starts reading the moment the plan is uploaded. By the time the designer presses submit, the hardest part is done. ✅
How do subagents split one job into many?
2️⃣ Split one job into many.
Checking five dimensions is five separate checks. Rendering the living room 🛋️, kitchen 🍳, bedroom 🛏️ and bathroom 🛁 is four separate images. One agent does them one by one. A lead agent can hand each one to its own subagent 👥, and they all run at once.
Four renders: 321 seconds one after another 🐢. 82 seconds side by side 🐇.
Can every step run at the same time?
🔗 Not everything can be done concurrently.
You can't render a room before it's built, or build it before it's laid out. The steps keep their order. Subagents only work in parallel inside a step, and they all read from the same plan, so nothing has to be pieced back together afterwards. 🧩
What should you ask when you build an agentic workflow?
💡 So the next time you build an Agentic Workflow, ask 2 questions:
- What can start immediately without losing speed or context? ⏩
- Which steps are really several jobs that don't depend on each other? 🔀
⚡ Same work. Half the wait or more.
👇 Below is the actual time spent on a real case.
If you are thinking about building an AI workflow or solving a business need, WhatsApp me or email me.
About the author
Kenneth Phua
Founder, Konverta.ai
Kenneth is a Singapore-based digital, data and AI leader with 15+ years of experience, including senior global roles at Coca-Cola, where he was the Global Head of Digital Strategy and Innovation. He has advised C-suite leaders across FMCG, retail, property, hospitality and luxury, and now applies the same playbook to Singaporean and Asian corporates and SMEs through Konverta.ai.
The newsletter
Weekly insights in digital, data and AI.
From bare floor plan to photoreal render in an hour
Where IDstudio started: with CAsarté Interior, an AI design engine that goes from a bare 2D floor plan to photoreal renders, 24x faster.
Next steps
Build with Konverta
Thinking about building an AI workflow?
We build business-critical AI workflows that run as fast as the work allows. See the AI design engine and the rest of what we build.
Prefer a form? Contact us.