What's new in DevEx - June 26

In a letter to the Senate, Anthropic accused operators tied to Alibaba's Qwen lab of running ~25,000 fake accounts and 28.8M Claude exchanges to clone its agentic coding skills. Millions of input-output pairs to train a cheaper copy that inherits the capabilities without the safety controls. High-volume API access is seemingly becoming a national security surface.

OpenAI previewed GPT-5.6 as three tiers: Sol ($5/$30), Terra ($2.50/$15), and Luna ($1/$6). Sol is promising a real jump on long-running coding and agentic work. But following a June executive order, it ships first to ~20 "trusted partners" named to the US government, via API and Codex only. After Fable 5 vanished by government order, that's twice in a month frontier access has been gated by Washington.

Sakana AI released Fugu and Fugu Ultra: a single model trained to call a pool of other LLMs, including recursive copies of itself, behind one OpenAI-compatible endpoint. It handles selection, delegation, verification, and synthesis internally, so the multi-agent complexity never touches your code. Sakana claims Ultra matches Fable 5 and Mythos.

AI SDK 7 introduces HarnessAgent, one API that runs full harnesses: Claude Code, Codex, Pi, and lets you swap them like you swap models. It normalizes the messy layer above the model call: skills, sandboxes, sessions, permissions, compaction, sub-agents. Switching agents used to mean rewriting your app around them. Now it's a config change.

With Wrangler 4.102+, an agent can run a single temporary-deploy command and push a live Worker for 60 minutes, with no browser, no OAuth, and no token, then hand a claim URL back to keep it. Agents even self-discover the flag from Wrangler's unauthenticated hint, so you don't hardcode it. It strips out the human-shaped friction that blocked agents from provisioning their own infra.

Oak is a content-addressed VCS where the unit of work is a branch, not a commit, and you never touch main locally. A Rust core with content-defined chunking and lazy mounts means no full-tree re-hashing and no clone wait, with p50 latency up to 95% lower than Git on snapshots and setup. A bet that branch-per-session belongs in the substrate, not bolted onto Git.

🌉 AI Engineer World's Fair is coming to SF next week!

The largest technical AI conference in the world: 6,000+ AI engineers, founders, and VPs of AI, 300 speakers across 29 tracks, and 100+ expo partners including Anthropic, OpenAI, Google DeepMind, Braintrust, Modal, and PostHog. Grab tickets here.

Till next time,