Why I Stopped Using One AI Agent for Everything
I fell in love with Windsurf for about three months.
The agentic coding experience was different. More fluid than Cursor. It felt like the assistant was in the project with you, not just answering questions about it.
Then I started building things complex enough to expose the problem.
Long sessions, the agent would lose the thread of what we were building and why. It solved the immediate problem and forgot the constraint from three problems back. Technically functional. Architecturally drifting.
The second issue was more subtle: one agent reviewing its own work is like one person writing code and signing off on their own PR. The review is happening. But nobody is checking.
In August 2025, I built PAIRED.
PAIRED is a suite of seven agents: PM, Architect, Developer, QA, UX, Scrum, Data. A coordinating PM agent routes the work. Each agent holds its own focused context. No one agent tries to be everything. The repo is at github.com/internexio/paired if you want to look at how it's structured.
The insight that came out of building it wasn't technical. It was organizational.
When you separate roles, you get natural review loops. QA's only job is to push back on Developer output. That's not a side task it has to remember to do. The Architect reviews structure without holding implementation details it doesn't own. You get adversarial depth from the structure, not from prompting an agent to "be more critical."
PAIRED also surfaced the memory management problem clearly. When a long session gets compacted, the plan in the early context evaporates. Single agents are fragile at session boundaries. Agents with bounded scope are stable across them.
I stopped using Windsurf eventually. Built my own environment. But PAIRED changed how I think about multi-agent systems, and that shaped everything I built after it.
I wrote about finding the second voice in the last post. PAIRED was where I found a version of it in code.
What would you put in a separate agent, if you could?