Your agent’s successful runs can become training data for a specialized model. LangChain launched LangSmith Fine Tuning yesterday, now in public beta. Its smithtune CLI takes curated agent sessions through dataset preparation, training with Fireworks or Baseten, evaluation and deployment. Those sessions include the messages, tool calls and results that led to an outcome. You’re teaching the model how to work through a task. In LangChain’s internal code review evaluation, fine tuning Qwen 3.8 27B improved F1 from 48.9% to 53.7% while using roughly 30% fewer model calls. What caught my attention: an earlier, less selective training dataset actually made performance worse. They improved the data selection and tried again. I think you should judge this on held out tasks and cost per successful run. Training on your own agent history sounds promising, provided you’re selective about which behavior you teach it.

One agent querying across AWS accounts without pulling every team’s data into one place? AWS has a reference design for that. Each team exposes a small set of MCP tools from its own account. AgentCore Gateway gives the agent one endpoint, and Cedar policies check the user’s claims before a tool call. Think a lending policy lookup without copying the policy library into the central agent account. Tool results still travel back to the central agent for model inference. Keeping source data put does not keep every answer inside its original account.

Hugging Face is bringing llama.cpp’s ggml kernels into Transformers. I like this direction for local AI. This walkthrough shows how a GGUF model loads, picks up compatible Metal kernels, and generates locally through the familiar Transformers API. The fast path currently targets Apple Silicon and Qwen3.5-compatible models. You’ll need Transformers from main and compatible kernels.

Before letting Hermes edit your project, turn on checkpoints. They’re off by default. Start your local CLI session with hermes chat --checkpoints. If a change goes sideways, use /rollback to list snapshots, /rollback diff 1 to inspect the changes, and /rollback 1 to restore that checkpoint. Enable it before you need it. Turning it on afterward won’t recover an earlier version.