The AI Journey So Far - From Traditional to Async Agent Teams

We've been through a few distinct phases already. Worth mapping out where we've come from and where the bottleneck is now.

Phase 0: Traditional development

For years, we wrote code the same way. Open the editor, type everything out manually, copy-paste from Stack Overflow occasionally. The engineer did all the work. All of it.

Phase 1: Copilot arrives

GitHub Copilot landed and suddenly autocomplete actually worked. Engineers stayed in charge, but the tedious bits — boilerplate, common patterns, closing out obvious implementations — got handled automatically. Behaviour shifted slightly: people started writing comments first, letting autocomplete flesh them out. Still very much human-driven though.

Phase 2: Agents take the wheel

Then came proper AI agents. Tools like Cursor, Windsurf, Kilo. The engineer moved from driver to navigator. You'd describe what you wanted, the agent would implement it. Less time thinking about syntax, more time expressing intent. Better use of human brainpower.

Phase 3: The current bottleneck

Here's the problem. Work only progresses whilst you're present. If you're not actively managing the agent — reviewing its output, giving it the next instruction, correcting course — nothing happens.

To genuinely unlock productivity, you need to explore multiple ideas in parallel. Try three different approaches to a problem simultaneously. Spin up agents to handle different features whilst you focus elsewhere. But engineers don't have the headspace to actively manage multiple AI agents at once.

The workflow needs to change. We need async agent teams that can work independently, report back, and let you review outcomes rather than babysit every step.

That's the next shift. And it will be here sooner than we think.