Nvidia Buys Hugging Face: Open Source Just Became a Moat

Keenable launched publicly on August 25 with a $26M seed led by Accel. Its index covers more than 100B documents for low-latency agent retrieval.

Treat retrieval as a roadmap workstream: define what your agent must find, then design its search and memory layers around those requirements. Start Monday by writing retrieval requirements beside the agent requirements, rather than leaving search as an implementation detail.

On August 27, The Trade Desk began rolling out Kokai Zuma, adding a conversational Koa Assistant plus specialized agents for campaign creation, audience building, troubleshooting, and performance analysis. Early results reported a 32% improvement in CPA.

Put the conversation layer and the orchestration behind it on the core roadmap, with separate success measures for each job the agents handle. For adtech and enterprise tools, agent coverage is now a product surface to plan and evaluate, not an optional add-on.

This week, Brian Kuan released *halo-record*, an open-source Python package for tamper-evident agent logs. Its append-only, hash-chained records cover tool calls, model calls, data access, and approvals, with redacted summaries intended for compliance use.

Treat an independent audit layer as a baseline product feature for agents, not a compliance patch added after launch. Integrating verifiable records lets customers check what an agent did without taking the vendor's word for it, which is a fairly low bar for software allowed near consequential work.

An Aug. 30 VentureBeat analysis argues that agents need distinct identities and delegation context, not merely successful authentication, before they execute autonomously. A companion analysis points to post-authentication drift, data exposure, and memory poisoning, calling for task-scoped credentials and attributable telemetry.

Treat identity, delegation, and continuous runtime trust checks as product primitives, not infrastructure fine print. The agent experience should make scope visible and support task-scoped credentials plus telemetry that ties each action to the agent responsible.

On Aug. 25, the Linux Foundation announced TRACE, an open specification for hardware-attested runtime and compliance evidence in AI agents. Supported by AMD, Intel, and Microsoft, TRACE creates portable artifacts that bind an agent's runtime environment, policies, and tool usage.

Add standardized attestation tooling to agent deployment evaluation, not as a later compliance patch. The resulting evidence gives governance and audit teams a cryptographically verifiable record of what the agent ran with and used.

This week, Google released Gemini Omni 1.1 Flash with reference video input, 4K upscaling, and continuing video generation. It also reportedly led human-preference evaluations for text-to-video quality.

Treat the model as a production tool, not the workflow: prioritize where reference inputs, generation, and 4K output fit into existing media journeys. The differentiator now moves to creative control layers, because quality scores alone don't specify how teams steer or revise a result.