At 33 tool calls the model lost short-term memory, coherence dropped to 58%. Adding KV-cache management and deterministic system prompt brought it to 87-99%.

What made the final workflow hit 99% tool success?

The full chain: MCP + RAG + KVCM + Layer Placement + Deterministic Prompt + Fine-Tuning Domain. Each added 5-15% quality in final weight. See How I Score above for pts (3/1/0) and 22 identical runs.

How long does one research run take?

About 51 minutes for 33 tool calls on consumer hardware (Gemma 26B via llama.cpp, local), slower than frontier API but with self-correction active and fully local.

Automation agent workflow for business research needs.