Efficient models, agent governance, and agentic security lead the last 24 hours. DeepSeek says its smallest new-architecture model beats its flagship at lower cost, Harness reports that enterprise confidence in agents outruns real controls, and Zscaler launched an agentic SOC for machine-speed threats. Video-training-data agents, a seven-framework field guide, and new workplace AI rules round out the day. Here are six stories worth your time, each summarized in our own words with a link to the original reporting.
DeepSeek ships V4.1-Flash and says it beats the flagship Pro
DeepSeek released V4.1-Flash, the smallest model in a new architecture family, and claims independent tests put it ahead of the much larger V4-Pro on performance, cost, speed, and total runtime. The mixture-of-experts design keeps only a small slice of parameters active per step, builds image understanding directly into the model, and sharply shrinks the key-value cache to cut memory and storage needs. Starting September 14, API calls to V4-Pro will be answered by V4.1-Flash at Flash pricing until a V4.1-Pro arrives, and weights are available under the MIT license.
Read the full story at SiliconANGLE (September 10, 2026).
Harness finds enterprise agent confidence outruns real controls
Harness published its State of Agent DLC 2026 report from a survey of 700 engineering leaders and found confidence in AI agents consistently ahead of the testing, security, inventory, and rollback controls that would back it up. Most respondents trust their agents but far fewer run active discovery, automated quality gates, or instant kill switches, and most report more production incidents since deploying agents. The report recommends treating the agent lifecycle as its own discipline with repeatable standards and progressive rollouts rather than routing agent changes through code-only pipelines.
Read the full story at Harness (September 10, 2026).
Zscaler launches Agentic SOC for machine-speed threats
Zscaler introduced Agentic SOC and the broader Agentic SecOps operating model to move security teams from reactive alert handling to proactive exposure reduction and machine-speed containment. The pitch combines large-scale inline telemetry with specialized agents for triage, investigation, and response, plus closed-loop enforcement through the existing zero-trust controls. Teams evaluating SOC automation should compare this telemetry-first pattern against their current stack of separate detection and response tools.
Read the full story at Zscaler (September 9, 2026).
Versos AI pairs agents with NVIDIA NeMo for video training data
Versos AI announced an agentic curation workflow built with NVIDIA NeMo that lets AI teams describe the video dataset they need in plain language and get back a structured, rights-cleared collection. The agents translate one natural-language request into structured criteria, search scene- and frame-level video intelligence, grade candidates against the specification, and assemble the dataset with ownership and provenance attached. The company will demonstrate the capability at IBC2026 in Amsterdam from September 11 to 14.
Read the full story at MarTech Series (September 10, 2026).
Kanerika compares the seven agent frameworks worth evaluating
Kanerika published a detailed September 2026 field guide covering LangGraph, CrewAI, Microsoft Agent Framework, OpenAI Agents SDK, Google ADK, Pydantic AI, and Strands Agents, with versions verified against package registries on September 10. The guide notes AutoGen is now in maintenance mode with Microsoft Agent Framework as its successor, and argues selection should turn on durable execution, mid-run approvals, memory, evaluation, and native MCP and A2A protocol support. Practical defaults include LangGraph for stateful workflows with approvals, CrewAI for role-based teams, and a cloud-vendor SDK when identity and compliance must stay inside one boundary.
Read the full story at Kanerika (September 11, 2026).
Workplace AI rules tighten across Colorado, California, and Connecticut
Epstein Becker Green summarized a busy first half of 2026 for employers using AI in hiring and workforce decisions, with several state laws effective since January 1 and more deadlines ahead. Colorado replaced its original AI act with a narrower notice-focused law effective January 2027, California's No Robo Bosses bill awaits a September 30 decision, and Connecticut requires disclosure when AI contributed to certain mass layoffs starting October 1, 2026. The authors also flag hiring-bias litigation testing whether vendors and deployers share liability, and recommend tool inventories, notice protocols, human checkpoints, and vendor-contract reviews.
Read the full story at National Law Review (September 1, 2026).