loc bengaluru, ist | local --:-- srijanshukla18@gmail.com →
[post]/tech/model-overthinking-is-a-control-problem

high thinking budgets make agents harder to control

/ 3 min read· ai

High thinking budgets can reduce control, especially when stale context and autonomous tool use mix together.

People usually assume more reasoning tokens = better results and more control. Lately I’ve been running into the exact opposite.

Give a model a huge thinking budget (high / xhigh) and watch the reasoning trace. Half the time it goes completely off the rails - inventing side quests, overcomplicating simple tasks, doing random speculative checks you never asked for. Even when your prompt is super strict! It’s not ignoring your instruction, it just thinks for so long that it talks itself into a much bigger scope.

When I actually want the model to do only what I asked and nothing else, setting thinking to low works way better. Less room for it to sit there and invent a whole new mission for itself.

Where this really bites you is in autonomous agents like Hermes, OpenClaw, or Pi where the model has tool access and runs without asking permission on every step. In a normal chat UI, if the model overthinks, you just get a long-winded answer. In an agent loop, overthinking turns into rogue tool calls, editing files it shouldn’t touch, or going down a 10-step rabbit hole that looks “reasonable” in its head.

It gets 10x worse if you reuse the same long session across multiple tasks. Now you have stale context from 20 minutes ago sitting in the window plus a huge thinking budget trying to connect dots between your old task and your new prompt.

You don’t notice this as much in default Claude Code or Codex coz they pause and ask for confirmation before running sketchy commands, so you can just hit No when it starts doing too much. With background/autonomous agents, you don’t get that luxury.

So yeah, cranking thinking to max is not a free upgrade. If you want predictable execution: keep the context clean, keep the prompt tight, and dial the thinking budget down.

Maybe Marc Andreessen was right about retardmaxxing.