ggerganov's "inception" pattern: inject a thought to force agents to act when they reason too long
olcan · x · 2026-08-17
Georgi Gerganov (llama.cpp) shares an inception pattern useful in agentic workflows: when the model thinks for too long, you inject a thought to force it to take action.
Implementation via limiting max reasoning length:
--reasoning-budget 4096 \
--reasoning-budget-message "... I am thinking for too long -- let me gather more info about the task."
Once the reasoning budget is hit, this message interrupts the model's thinking and pushes it to gather information instead. Gerganov notes it's especially helpful for underspecified tasks that make models reason endlessly. Tune the values to your needs.
More from coding & agent
- When a Template Grows Up: My Obsidian Publish Workflow Became a Full Plugin — dSebastien · 2026-08-17
- Open-Sourced 'One Agent, Many Hats': An LLM Agent That Learns on Its Own — sandeepkavety · 2026-08-17
- bloub: Open-source SVG recreation of the x.ai bot avatar morphing through 14 states — iannuttall · 2026-08-17
- POLYBOT: An Open Source UI Middleware for Orchestrating Multiple AI Agents — RileyRalmuto · 2026-08-17
- olore: A Package Manager for Docs That Feeds AI Coding Agents Locally, Offline — tom_doerr · 2026-08-17
- Viral Codex skill turns photos into abstract editorial posters, hits 2.8K GitHub stars — oran_ge · 2026-08-17