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:
- a thin planner keeps only global state and routing
- executor steps get small task briefs and return results
- the planner keeps only the decision-relevant summary, not the full worker context
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.
More from coding & agent
- Previewed human-plus-AI workspace defaults to agent replies and shared cloud spaces — Scobleizer · 2026-07-28
- Survey of 126 respondents says company-brain hype has moved beyond X — femke_plantinga · 2026-07-28
- Unifying Outputs from Thousands of Parallel Agents in Forecasting — they_will · 2026-07-28
- Ycode adds AI agents that work inside its website builder with Claude, ChatGPT, Gemini, or Grok — JaynitMakwana · 2026-07-28
- A Claude agent meme turns prompt instructions into pure workplace abuse — pranavmarla · 2026-07-28
- Study of 25,264 agent PRs says coding agents may not improve team collaboration — Suspicious_Orchid770 · 2026-07-28