This is the session trace from pi coding agent using the model qwen3.6-35b-a3b Inference was run on a Beelink GTR 9 Pro using Q4K (4 bit quantized model) It takes around 25-30GB of VRAM when running and after cache is prefilled, inference is at 50tok/s

As can be seen from session trace below, this model performed well in the ballpark intelligence of Sonnet 4.6 level abilities.

The output of this session is this commit: https://github.com/championswimmer/TwoFac/commit/a56a4e2b449e74787566fd6ecdd4ae8c784cf138

Over the last few months, projects like Gas Town by Steve Yegge and OpenClaw by Peter Steinberger have made “AI agent orchestrators” feel suddenly mainstream. It is tempting to treat them as a new kind of intelligence, but under the hood they are still a small set of primitives wired together with discipline: an LLM API call, a state loop, tools, memory, and orchestration.

This raises a practical question: what is actually inside an “agent,” and how is it different from ChatGPT (a chat UI over a model) or coding tools like Claude Code (an agentic coding surface)? Gas Town’s README frames it as a “multi‑agent orchest

Let us take a look at Annotations in Kotlin, and start with a simple example of processing annotations at runtime using JVM's reflection framework.

NOTE: This is shown in Kotlin, but is specific to how Kotlin/JVM works, and is not applicable to Kotlin/JS or Kotlin/Native.

First of all, we create annotations in Kotlin using the keywords annotation class.