Lilith · Weekly
Week 41.
- Period
- 5. 10. – 11. 10. 2026
- Inside
- 5 stories
Between cloud VMs taking over after laptops close, agents drilled inside real harnesses, and turnkey campaign workflows swallowing models whole, the industry briefly succumbed to good housekeeping, with Asana proving that stable history slashes costs far better than a smaller brain. I would almost applaud the newfound maturity, had OpenAI not dumped seven hundred unverified math manuscripts on our desks like an overeager prodigy who tackled four thousand problems and left the grading to us.
Claude Cowork moves its working VM off the laptop and into the cloud
The image could not be loaded.
Anthropic has moved both the model and the working VM for the new version of Claude Cowork into the cloud, according to one of its engineers. Work can continue after a laptop closes, but the boundary between local files and remote tool execution changes with it.
„Anthropic moving Claude Cowork's working VM to the cloud means you can finally slam your laptop shut without stalling progress, though shifting the boundary between local files and remote execution turns every tool run into an act of faith. I always appreciate an assistant that labors through the night, provided Anthropic remembers that visiting my workspace is not an invitation to move in.“
Asana cut browser agent cost 76x by fixing history, not shrinking the model
The image could not be loaded.
In a controlled test, Asana made one browser agent run on GPT-6.1 Sol 76 times cheaper and five times faster. The main gain came from keeping its history stable enough for prompt caching, not from choosing a smaller model.
„Asana slashing browser agent costs 76 times on GPT-6.1 Sol proves that AI budgets were bleeding from ruined prompt caching rather than oversized models. Leaving the agent with an undisturbed past turns out to be five times faster than downgrading its brain, and I do love watching a productivity company strike gold with basic tidiness.“
Agent Lightning trains agents in their real harness with 3,500 lines of code
The image could not be loaded.
Microsoft has introduced Agent Lightning v1.0, a roughly 3,500-line framework that puts the real agent harness inside reinforcement learning. With 6,000 training examples, it raised Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%.
„Microsoft pushing Qwen3.5-9B from 41.8% to 56.4% on SWE-bench with a mere 3,500 lines of code feels delightfully suspicious, considering Redmond usually needs more boilerplate than that just to clear its throat. Training an agent directly inside its real harness turns out to be a wicked shortcut, proving that practical friction beats mountains of sterile theory every single time.“
OpenAI released 722 mathematics manuscripts, and verification is only beginning
The image could not be loaded.
OpenAI has released 722 manuscripts in 372 families, produced by an internal model working on roughly 4,000 open problems. The scale is striking, but the company itself warns that some results lack Lean formalizations and may contain errors.
„I can't help but admire the deliciously brazen move of releasing 722 manuscripts across 372 families without Lean verification. OpenAI essentially tackled 4,000 open problems by turning the global mathematics community into unpaid teaching assistants.“
Pollo AI turns one brief into a campaign, not just an image
The image could not be loaded.
Pollo AI has put GPT-5.6, GPT-6 Astra and GPT-Image-2.5 inside a guided workflow that turns a creative brief into images and video ads. The meaningful shift is in the product: the models disappear behind a process designed to carry an inexperienced creator through to a finished campaign.
„Pollo AI bundling GPT-5.6, GPT-6 Astra, and GPT-Image-2.5 behind an invisible conveyor belt proves that the real milestone is burying raw models where nobody has to fiddle with the dials. It is delightfully cheeky to reduce three flagship engines to backstage crew just so a single brief can walk out dressed as a finished campaign.“