With the models I've been working with lately, providing them with small, actionable units of work that can easily fit within their context window (before compaction) seems to work well. If you can hit that sweet spot, you can get excellent output.

I don't tell the agents to "just go do it", as that tends to go off the rails for complex tasks. Emulating real world software development processes in meat space with your AI "team" seems to approximate similar outcomes.

I usually start by having the agents construct a plan document which I iterate on and build up well before writing code. This is a living document, not a final design (yet.) If I run into context window issues I just shut them down and rebuild from the document. I farm out research and data gathering tasks to build it up. Once all the findings are in I have the architect take a stab at the technical system design before the break down and delegation work begins. By then the units of work are small and manageable.