My approach is simple.
I keep multiple sessions running. Each session is an agent with a specific persona and a specific goal.
Think of a startup with three or four team members. One session acts as the CEO. Another is the planner. Another is the designer. Each is responsible for a different part of the outcome.
Why separate sessions work
This keeps you from overwhelming a single agent. Each session carries less irrelevant context, which saves cost and reduces context bloat and compaction risk.
The agents share the same codebase and work through committed changes. Git history becomes the traceability layer: it shows what changed, when it changed, and how the work evolved.
Add an advisor
You can add another persona: an advisor agent that checks in from time to time. Its job is not to build a feature. Its job is to ask how, why, and where the startup is drifting.
Focus each agent on one goal. Use the shared history to keep the whole team aligned.
The structure is closer to a small team than a single, endlessly growing chat session.
View the original post on X ↗