What agents can now output, how to vet agents you deploy, and where the document-AI procurement bar just moved.
Agents that produce video, not just text (workflow).
HeyGen open-sourced HyperFrames, a library where HTML and CSS templates render directly to video. The structural change for operators scoping agent workflows: video output is now a realistic deliverable format for any agent that already produces HTML or structured data. Explainer clips, data visualizations, status reports in video form: the video encoding is handled by the renderer; the agent writes what it already produces. The caveat worth scoping: what the video says still matters as much as it ever did. Cheaper production changes the cost floor, not the content bar.
Prompt injection is a trust problem, not a filter problem (procurement).
The root mechanism, from a new ICML 2026 paper by Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell: agents need to track which instruction came from which trusted principal, and current LLMs identify roles through stylistic features rather than formal trust boundaries. Simon Willison flagged the paper for the reframe it makes concrete. User messages, tool results, system prompts, and malicious content in retrieved data all arrive as text; what the architecture has to enforce is which source the agent treats as authoritative. The procurement implication follows: input sanitization addresses a narrower attack surface than this requires. The meaningful question for any agent platform evaluation is whether the architecture enforces instruction provenance at the trust boundary, and that distinction is worth pressing before you sign.
Long-document OCR baseline just moved (procurement).
Baidu open-sourced Unlimited-OCR, a model that parses arbitrarily long documents in a single pass, without chunking pages into sections and stitching results back together. The stitching step is where most OCR pipelines accumulate errors on long contracts, filings, or research documents. For anyone procuring or re-evaluating document processing tools, the incumbent chunking-and-stitching architecture is no longer the only option. There is now a concrete open-source benchmark for asking whether a document processing vendor carries the same stitching limitation, and what that costs in accuracy on a 50-page document.