27B Beats 75B in Agentic Tasks
Important_Quote_1180 · reddit · 2026-07-10
The author tested two models on agent tasks: an untuned 27B completed all tasks in 6-9 tool calls under a neutral system prompt, while the 75B required manual prompt tuning to barely pass, with higher call counts and latency.
The post emphasizes that for agents, "fewer turns" matters more than "faster tokens." It also details a caching-related failure case: repeated runs with the same payload failed consecutively, likely due to prefix caching issues, which was later bypassed by adding a nonce to the system prompt.
More from coding & agent
- Codex helps build Valdiluce, an open-world game with climbing, gliding and gondolas — Dimillian · 2026-07-22
- HeyGen adds a media-sourcing skill for coding agents with 75k images and 10k tracks — HeyGen · 2026-07-22
- Agent search bottlenecks are now about variance, not raw latency — rohanpaul_ai · 2026-07-22
- LangSmith adds tracing for Pipecat, LiveKit, OpenAI Realtime, and Gemini Live — LangChain · 2026-07-22
- An MCP server signs every AI agent tool call into a verifiable Merkle chain — Funky_Chicken_22 · 2026-07-22
- Annotated transcript of a Claude Code team interview is now available — trq212 · 2026-07-22