Designing Mac agents without Full Disk Access
Apple is tightening Full Disk Access on macOS. These are my field notes on building scheduled Mac agents that do not depend on that permission.
What AMD buying World Labs actually signals
AMD buying World Labs matters less as an acquisition story than as a signal about persistent 3D worlds, future AI workloads and compute design.
What a fixed API shutdown date should change
A practical studio process for fixed API shutdowns: isolate the call, own your inputs and outputs, and treat a missing successor as procurement.
What PixVerse R2 changes in a video pipeline
PixVerse R2 treats video as a persistent session, not a stack of clips. Here is what that changes for real client pipelines, and what I would test first.
Qwen-Image-2.1: the licence comes first
Qwen-Image-2.1 adds native transparency and ten reference images, but its research-only licence keeps it out of client pipelines without a separate deal.
Designing for model churn
Higgsfield’s new API and Kling’s endpoint retirements are a reminder to build video pipelines for change, not for one model or one vendor alone.
Who should own your agent loop?
OpenAI’s managed Agents API removes runtime plumbing, but for client work the harder choice is who owns state, data paths and operational control.
Cognition’s SWE-2 and where coding agents are heading
SWE-2 points to a new coding-agent trade-off: frontier-level eval performance, specialised RL training and materially lower cost for software work.
GPT-6 Astra still needs boring engineering
GPT-6 Astra raises the ceiling for agentic computer use, but reliable client automation still depends on verification, tight permissions and cost control.
Read the post →