← Journal

May 22, 2026

Two releases worth your attention today. LlamaFactory v0.9.4 raises its Python floor to 3.11, swaps in uv for package management, and adds Orthogonal Fine-Tuning alongside Megatron-LM training support. If you are still on Python 3.9 or 3.10, migration is now a prerequisite. On the ops side, Vercel's CLI gains a `vercel alerts` command that surfaces anomaly alerts and AI investigation results inline, cutting out the browser loop for developers and agents alike.

research
LLMs Can Now Think, Read, and Act at the Same Time

A new paper proposes replacing the single-stream message format that powers today's AI agents with parallel computation streams, letting models read, think, and generate output simultaneously. The change is data-driven and delivered through instruction-tuning on multi-stream formats.

coding_agent
One Sandboxed Runtime to Run Every Coding Agent Your Team Uses

Runtime gives teams a single shared environment for running coding agents like Claude Code, Cursor, Codex, Copilot, and others, with built-in governance, cost tracking, and real-time collaboration. It is part of Y Combinator.

ops
Vercel CLI Now Surfaces Anomaly Alerts and AI Investigation Results

Vercel engineers can now pull anomaly alert details directly from the CLI using the `vercel alerts` command. AI investigation results are also available inline, letting developers and agents act without leaving the terminal.

ops
Dify Patches Tenant Isolation, Workflow Tracing, and Tool Credentials

Dify v1.14.2 tightens tenant isolation, fixes workflow tracing after human-in-the-loop resume, and locks down tool credential management. Here is what changed and what to check before your next deploy.

framework
ByteDance Rewrites DeerFlow 2.0 as a Full Super Agent Harness

DeerFlow 2.0 is a ground-up rewrite from ByteDance that orchestrates sub-agents, memory, and sandboxes through extensible skills. It hit the number one spot on GitHub Trending on February 28th, 2026, the day version 2 launched.

research
Turn Agent Trajectories Into Long-Context Training Data

Agent Context Compilation (ACC) converts multi-turn agent trajectories into long-context QA pairs, giving smaller models a cheap path to long-range reasoning without expensive document curation. Qwen3-30B-A3B trained with ACC matches Qwen3-235B-A22B on key benchmarks.

framework
LlamaFactory Drops Python 3.9, Gains OFT and Megatron Support

LlamaFactory v0.9.4 raises its Python floor to 3.11, migrates to uv for package management, and ships Orthogonal Fine-Tuning, Megatron-LM training, and over a dozen new model integrations. Here is what builders need to act on now.

framework
Agency Lives in the Model. Your Job Is Building the Vehicle.

A new open-source curriculum reframes how engineers should think about agent products: the model supplies agency through training, and the harness is the infrastructure that lets it operate. Here is what that split means for builders.