What I learned building a multi-agent platform people can use entirely from their phone
For the last few weeks I've been building AtiqSoft Agents (agents.atiqsoft.com), a platform where 290+ specialist agents work in teams, with a lead agent routing tasks to the right specialist. The goal is to remove PC dependency: from a phone chat, the agent can build apps in a private cloud dev workspace, create Word, Excel, PowerPoint and PDF files (including Arabic and Urdu), and show live previews.
A few things that surprised me:
Cost leaks hide in agent loops. One hardcoded model choice in the coding agent billed every step at the premium tier, even when users picked the cheap one. Now every step follows the user's chosen tier, free models go first, and there's a daily safety cap.
"Done" means nothing without verification. The agent used to say "done" on a page with no CSS loaded. Now it screenshots its own preview at phone width and compares it to the reference image before finishing.
Free models are good at chat and weak at multi-step coding. They break the step format more often, so retrying on a different free model works better than failing.
Parallel tasks need isolation. Each task gets its own folder and port, or they fight over the same preview.…
