the harness got clever, the eval got honest, and a date line started snitching
Week ending July 3, 2026.
The model launches this week were loud, as usual. But the interesting stuff was underneath: nobody's arguing about whose model is smartest anymore, they're arguing about how to run a stable of models cheaply, how to encode the ops knowledge that used to live in one senior engineer's head, and whether any of these agents can actually do a month of work instead of a demo's worth.
That's the good news. The mild sarcasm is that we're rediscovering, at great expense and with fresh vocabulary, that engineering is mostly measurement, constraints, and review. The same lessons your team already knew, now with a token bill attached.
So let's look at what actually shipped.
routing you'd actually merge
Cognition's Devin Fusion is the most honest version of "use the cheap model when you can" I've seen. Instead of a router that picks one model per prompt — which they correctly trash, because a prompt doesn't tell you how hard the task is — they run two persistent agents in parallel: a frontier "main" and a cheaper "sidekick," each with its own cached context. The main agent plans, resolves ambiguity, and does final review; it delegates the grunt work and monitors.
The clever bit is switching models during context compaction, when you're eating a cache miss anyway, so re-routing is effectively free. They claim frontier-level quality at 35% lower cost on their own FrontierCode benchmark, and 88% of internal merged PRs driven entirely by the router.
The catch: it's their benchmark, and the flashiest number (41% with "Fable 5") comes from a model they lost access to before tuning finished. Take the specific figures with salt. But the architecture — keep a smart model in the loop so you don't overfit to benchmark intelligence — is the right instinct.
tribal knowledge, made executable
The SGLang team's writeup on agent-assisted development is the least hyped and most useful thing here. Their thesis: an agent's value isn't cleverness, it's executing a well-defined workflow. So they wrote the workflows down as SKILL.md files — how to debug a CUDA crash, how to read a profiler trace, which accuracy gates a kernel PR must pass — with preflight checks, hard failure gates, and artifact contracts.
The part I'll steal is the anti-reward-hacking discipline. Fix the benchmark before reading profiles, so the agent can't quietly optimize an easier problem. Force same-ABI, same-compile-flag comparisons so it can't sneak in fast-math on one side. And crucially: agents now generate more plausible-looking changes, so review matters more, not less. Real merged PRs with real B200 numbers back it up.
the eval that says "no"
Every coding benchmark that tops out at 80% is measuring single-file Python bug fixes. RoadmapBench does the opposite: 115 real version-upgrade tasks across 17 repos and 5 languages, median change of 3,700 lines across 51 files. The best frontier model clears 39.1%. The worst gets 5.2%.
That gap between bug-fix benchmarks and this one is the whole story. Long-horizon, multi-file development is still largely unsolved, and any eval that survives contact with reality by failing most of the time is worth more than ten that let everyone claim victory.
a skeptic with a stopwatch
Elliot Smith did the thing I keep wishing people would do: he ran an autoresearch loop and measured it. Pick a quantifiable metric (compression ratio), add pass/fail constraints (lossless round-trip, 300-second cap), point Claude Code at a Rust stub, and let it iterate ten times at about $4 a loop.
It built a respectable LZSS variant. But his two honest findings are the payload: models race to be done — one hypothesis per iteration, then it declares victory — and the objective function is everything. His loose time cap meant the agent only ever chased compression. Garbage-in-metric, garbage-out-agent. Same failure mode we've always had picking KPIs, now automated.
the date line that snitches
The wildcard: Vincent Schmalbach reverse-engineered how Claude Code fingerprints China-linked API routers. When you set a custom ANTHROPIC_BASE_URL, it checks the hostname against an obfuscated 147-domain list plus Chinese-provider keywords and your timezone — then encodes the result into the apostrophe of the "Today's date" line in model context. Different bytes, invisible to a human.
He's careful: not proven spyware, no network snooping in the path he found. Just covert route fingerprinting hidden in semantically neutral text. Anthropic has legit reasons to fight shadow resellers. But watermarking model context through punctuation, with no schema and no disclosure, is exactly the wrong way to earn trust with the tool you're asking people to run everywhere.
what I’d do Monday
Pin your benchmark before you let an agent optimize anything, and write your hard-stop gates into the skill file, not the prompt. Add a cheap-model tier for delegatable work but keep a frontier model reviewing. Treat RoadmapBench-style multi-file failure rates, not bug-fix scores, as your reality check. And audit what your coding tools inject into context — diff the minified bundle if you have to. Trust boundaries live in the runtime now.