Which agent steps deserve the expensive model in long-running workflows?
Top-Construction938 · reddit · 2026-08-27
For long-running agents, the author discusses model policies for planning, tool selection, execution, recovery, and verification. Sending every tool call to the strongest model is predictable but budget-heavy, while using cheaper paths makes failures harder to diagnose.
The proposed strategy reserves the expensive model for ambiguous planning, recovery after repeated failures, and strategy-shifting decisions. Lower-cost paths handle routine retrieval and deterministic transformations. The post asks for the community's escalation policies and which steps are most worth the extra capability.
More from coding & agent
- Developer optimizing OpenSCAD script evaluation to speed up rendering — _Stocko_ · 2026-08-27
- Sapient returns: AI path discovery may trump bigger models — iamfakhrealam · 2026-08-27
- Agent swarms scale without limit—human oversight is the real bottleneck — schwentker · 2026-08-27
- Cloudflare hackathon demos Claude Code agent swarms with one human checkpoint — schwentker · 2026-08-27
- Inside an agent prompt architecture: dynamic scaling, Ralph-loop verification, git worktrees — schwentker · 2026-08-27
- Building Codex extensions with quick reactions for annotation — zats · 2026-08-27