Heavy AI Users May Waste More Capacity on Routing Than on Prompting
Druss_ · reddit · 2026-09-25
Running a multi-agent system for real work, the author found that locking one model and reasoning level for an entire session—planning, research, trivial checks, hard reasoning, validation—is inefficient. Decomposing work helps: deterministic steps use no model, cheap steps go to fast models, hard steps get strong reasoning, validation is separate, and failures escalate instead of paying max cost upfront. Several small-model runs consumed visibly less paid capacity than one large manual run. His thesis: treat model selection as resource allocation, like assigning tasks to the cheapest competent employee level and escalating exceptions.
More from coding & agent
- Beyond embeddings: grep agents and whole-corpus long-context prompts as underrated retrieval patterns — jakedahn · 2026-09-25
- Nokia Open-Sources AnyJev: Turn Any LLM into a Calibrated Decision Model, No Training — kalyan_kpl · 2026-09-25
- Microsoft's biggest Copilot update ever: Autopilot agents, Code, Home, and embedded Office — satyanadella · 2026-09-25
- TSP: Self-Hosted LLM-Powered A-Share Quant Workbench Passes 5,000 GitHub Stars — shy3130 · 2026-09-25
- Starnet: Pixel-Art Desktop Harness for Watching AI Agents Work, Built Local-First — androoAGI · 2026-09-25
- huashu-chrome: open-source MCP server lets AI agents drive your logged-in Chrome with 22 tools — AlchainHust · 2026-09-25