Coding agents get a long-context model and a critique of GitHub

Agents strain the tools around them

Moonshot's model documentation for Kimi K3-256k specifies a 256k token context window aimed at coding and agentic use. As an open-weight release, it gives engineers a concrete model to point coding agents at now, though a context window that large may raise hardware questions for teams that want to run it themselves.

Depot's post addresses what happens after agents produce code. It argues that GitHub's core interaction model, built around pull requests and human review cadence, increasingly fits poorly when AI agents generate and revise code much faster than people review it. That is an argument, not a measured finding. Still, it identifies a real friction point for teams adopting agents: the review loop was designed around human authors, and longer-context models aimed at agentic coding are unlikely to slow the pace at which changes arrive.

Recompiling RISC-V for faster emulation

For readers working closer to the hardware, a detailed technical writeup walks through building a RISC-V recompiler to push emulation performance beyond what typical interpreter-based approaches reach. The post focuses on recompiler techniques rather than a product. It should be useful to anyone building emulators, JITs or instruction-set translation layers, where the trade-offs between interpreting instructions and translating them ahead of execution shape overall speed.