Telling an agent it has 1M tokens vs 10K makes reasoning 3-5x longer, same model
paraschopra · x · 2026-09-10
paraschopra ran a small but telling experiment: agents told their token budget in the prompt behave very differently even on the same underlying model.
- The agent told it had a 10K token budget became frugal — switching to bash and grep to save tokens ("Be efficient. Let me use bash curl for API calls.")
- The agent told it had a 1M budget explored widely, reasoning "tokens burn slowly" and "plenty of runway", and produced 3-5x longer reasoning on the same task
The author believes this is not a sampling error and is visible in the reasoning traces, suggesting models have meta-cognitive abstractions and adjust strategy based on perceived resource constraints.
Related event: Telling an Agent It Has 1M Tokens Multiplies Its Reasoning 3-5x(2 posts)→
More from coding & agent
- Qwen3-VL-2B fine-tuned into a computer-use agent that does cell segmentation in a headless labeling app — secemp9 · 2026-09-10
- Prerender Buddy MCP server lets coding assistants audit crawler accessibility — kopachl · 2026-09-10
- Perspective AI turns static forms into adaptive AI conversations via MCP — modelcontextprotocol · 2026-09-10
- OilPriceAPI ships MCP server: query 40+ energy commodities via natural language — modelcontextprotocol · 2026-09-10
- Should half-built agents touch your real Gmail? Devs weigh test-account pain — aofu_dev · 2026-09-10
- QuantMind: open-source agent-native framework turns papers, news and filings into structured quant finance knowledge — tom_doerr · 2026-09-10