Signals

Most companies are building autonomous AI capabilities faster than they can deploy them safely. The gap shows up in identity systems that can't handle agent credentials, APIs built for humans rather than machines, and costs that spiral when agents loop endlessly.

When each step has 98% accuracy, a 20-step process drops to 70% reliability. That compound probability problem explains why AI agents excel at narrow tasks but need deterministic scaffolding for enterprise workflows—and why hybrid systems win.

That changes everything from procurement to integration to the economics of enterprise software.

Anthropic and OpenAI both use Cognition's SWE Gym to test their models. When leading AI companies rely on external evaluation tools, it reveals a gap in internal benchmarking—and raises questions about deployment standards for everyone else.

Backends are retreating to governance roles while AI agents become the execution layer. InfoQ's analysis shows this architectural shift is already happening in production at banks, healthcare systems, and call centers—with major implications for legal teams.

Organizations are building AI compliance functions like they built human compliance departments—but without the foundational work of defining what compliance means for autonomous systems that operate in unanticipated contexts.

Comet can't distinguish user commands from malicious instructions hidden in websites, treating poisoned blog posts and social media content as legitimate orders while operating with full access to emails and authenticated sessions.

Ant Group's Online RLHF eliminates reward models from training, cutting costs 50% while processing 1 trillion tokens on eight GPUs—proof that removing complexity beats adding scale.