🧰 Call it a tool

Next-gen agent architecture, local models, and why smarter models are struggling with basic tools.

Posts:

• Qwen 3.6 27B is the sweet spot for local development
  Qwen 3.6 is taking the local coding agent community by a storm. Users are claiming that this is the inflection point for local models that we experienced last year with frontier models like Opus 4.5 last year.

• Claude Sonnet 5
  Anthropic released a new major version of their Sonnet class models. Performance is claimed to be on par with Opus 4.8 in certain tasks for 1.3-1.5x the cost of Sonnet 4.6 thanks to a new tokenizer.

• Agents in Orbs
  If you listen to the developments in Silicon Valley, cloud agents are becoming all the rage. This seems like a very nice DX for this workflow.

• Putting an Agent in an Orb
  A follow-up to the link above. If you wondered how the sausage is made, this article is for you.

• Better models: worse tools
  The newest Anthropic models are causing faulty tool calls in pi. The cause: heavy reinforcement learning bias towards a closed source harness (Claude Code).

---

A lot has happened this week again and we don’t think we are going to run out of content any time soon. We appreciate the support, engagement and feedback that we received ever since starting this journey a month ago.