The improvements arrived outside the weights

Everything that got better this week got better outside the model. Microsoft open sourced a tool that upgrades an agent's skills without retraining anything, OpenAI shipped record and replay for Codex runs so a failed run can be played back rather than guessed at, and a practitioner post made the loop itself the unit of design. Even the storage win landed after the fact, on vectors a team already had. We wrote about this a few weeks ago from the other side: the harness was moving the numbers while the model stayed swappable. The harness, if that word is new, is everything wrapped around the model: the instructions, the tools, the loop. This week the harness got the tooling that comes with being a real layer: an optimizer, a debugger and a hosting story. Teams can now improve the parts they own, which is a nicer place to stand than the queue for the next checkpoint. The tell will be whether skills and loops end up in code review like everything else, or stay in prompts nobody diffs.

Top Stories#

Agent skills got an optimizer that never touches the weights. Microsoft released SkillOpt, which improves the skills an agent uses without modifying model weights. A skill here is a written instruction file the agent loads for a particular job, and what matters is where the improvement lives. A better skill is a file in a repository, so it survives a model swap, a person can review it, and it needs none of the training infrastructure almost nobody outside a lab has. Fine tuning, meaning retrain the model on your own examples, has been the assumed answer whenever an agent underperforms on a specific job. Optimizing the instructions instead puts that work back within reach of the team that actually has the domain knowledge.

The agent loop is becoming something teams can debug. LangChain published an argument for treating the loop as the engineered artifact rather than the thing that happens between prompts. Meanwhile, OpenAI shipped record and replay for Codex, which captures a run and plays it back instead of asking an engineer to reproduce it by hope. Non-determinism, the fact that the same prompt can produce a different answer twice, has been the standard excuse for why agent failures never get root caused. The randomness is still there, but now somebody can go look at the actual run.

Cloudflare is collecting agent harnesses rather than backing one. The company added support for more agent harnesses, starting with Flue, on a platform that already runs the network underneath a lot of other people's products. Selling the place every harness runs, rather than betting on one, is the same shape as hosting runtimes instead of picking a winning language. It tends to be the position left standing whichever framework the year decides it likes.

A 49 percent cut in vector storage arrived after the fact. A team reported the result, from a post hoc iterative residual method, applied to embeddings that already existed. Embeddings are the number arrays a search index keeps for every chunk of text, and they pile up. No re-embedding, no new model, no migrating the pipeline that produced them. Storage is the cost line that grows quietly with every document added, and it has mostly been treated as a fact of life. Compression that works on vectors already written suggests the index is full of slack nobody bothered to squeeze.

Top Links#

• Learn about Lore: next-generation open source version control (lore.org): a new open source version control system, which is not a category that sees many entrants.
• Unsloth pricing (unsloth.ai): the fine tuning toolkit puts prices on the page, worth reading as a signal about where that market sits.
• How the Open Knowledge Format can improve data sharing (cloud.google.com): Google's pitch for a shared format for the knowledge layer agents read from.
• Midjourney Medical (midjourney.com): an image model company with a medical page, which raises more questions than it answers.