AI digest: Agent infrastructure gets serious
Major releases in agent infrastructure this week, plus OpenAI quietly kills Sora after six months.
The agent tooling space is heating up with some proper infrastructure releases, while OpenAI made a curious retreat.
Amazon drops A-Evolve to automate agent development
Amazon researchers released A-Evolve, calling it the “PyTorch moment” for agentic AI systems. The framework promises to replace manual agent tuning with automated state mutation and self-correction. Bold claim, but if it delivers on making agent development less of a black art, this could be the standardisation the space desperately needs.
Agent-Infra launches AIO Sandbox for proper agent execution
Running agent code safely has been a proper headache, and Agent-Infra’s new AIO Sandbox tackles this with an all-in-one runtime. It bundles browser, shell, shared filesystem, and MCP support into one package. This addresses the real bottleneck in agent deployment, which has shifted from model reasoning to having a decent execution environment.
Chroma’s Context-1 challenges the big context window trend
Chroma released Context-1, a 20B model focused on agentic search rather than massive context windows. Their argument is that stuffing millions of tokens into prompts is expensive and often pointless compared to smart retrieval. Makes sense, though it goes against the current “bigger context is always better” narrative.
OpenAI quietly kills Sora after six months
OpenAI shut down Sora just six months after public release, raising eyebrows about whether this was a data collection exercise. The timing feels odd given the hype around AI video, but maybe the compute costs didn’t justify the engagement. Either way, it’s a rare public retreat for OpenAI.