Splitting planner and executor roles cut my agent context bloat and token bills

truecakesnake · reddit · 2026-07-28

Splitting planning from execution cut agent context bloat and token spend

The author says their agents were failing because they kept stuffing whole codebases and tool logs into one giant context window. Repeated context was drowning out instructions, token bills were exploding, and the model kept losing the signal.

They switched to a planner/executor design:

That change reduced token usage mainly because the planner stopped rereading the same heavy context every loop. They also tested MiniMax M3 for executor calls, especially long-context code/log reading and synthesis, because low enough input cost makes repeated large executor contexts practical.

The post ends by asking how others decide what should stay in planner memory versus what should be passed to executors without losing decision context.

Original post →

More from coding & agent

coding & agent channel →