[ The Morning Build ]
4 min read 5 stories AI

The Morning Build for August 29, 2026: Open-weights, agent safety, and persistent agents

Today’s stories matter for engineers building and operating agentic systems: GLM-5.3 landed as an open-weight release on Hugging Face; Nutanix published a three-layer defense-in-depth architecture for agent security with Intel and Cisco integrations; LAION released the Big Video Dataset for research; Anthropic unveiled a Model Hardware Standard research preview to unify device drivers; and OpenAI is testing a ‘Persistent mode’ for its Codex agents.

GLM-5.3 published as open-weight model on Hugging Face with 300k–1M token context support

  • What happened: Zai-org published GLM-5.3 as an open-weight model on Hugging Face with deployment instructions for Transformers, vLLM, SGLang, and Docker Model Runner. The release specifies evaluation settings including context lengths up to 300,000 tokens and mentions evaluation runs using up to 1,000,000-token windows and max generation lengths up to 163,840 tokens.
  • Why it matters: Engineers can run and benchmark the exact model artifacts locally or on custom infrastructure and reproduce the publisher’s post-training gains versus GLM-5.2, which the page quantifies across many agent and coding benchmarks. The model’s published deployment paths (vLLM, SGLang, Ascend NPU support, and Docker runner) let teams validate latency, memory, and context-management strategies in their own stack.
  • Outlook: Community reproductions and benchmark runs reported on the model’s Hugging Face repo and in third-party evaluations using the author’s evaluation settings (300k–1M token contexts and the specified temperature/top_p/max_new_tokens) will be the first public check on the claimed post-training gains.

Sources: huggingface.co · techcrunch.com

Nutanix lays out a three-layer defense-in-depth architecture for agentic AI with Intel and Cisco integrations

  • What happened: VentureBeat published a sponsored piece from Nutanix that describes a defense-in-depth architecture for autonomous agents consisting of an infrastructure layer (hardware-rooted trust, confidential computing), a network layer (dynamic policy enforcement and microsegmentation), and a centralized control plane (agent permissions, tool access, runtime visibility). The article highlights integrations with Intel (hardware trust, Xeon 6 AMX), Cisco (secure fabric and integration with Cisco Secure AI Factory), and Nutanix Agent Gateway.
  • Why it matters: The architecture assigns distinct responsibilities to hardware, network, and control-plane components so security controls scale with agents’ runtime privileges rather than duplicating the same app-level guards. Engineers planning production agent deployments get a concrete mapping of risks to controls: platform attestation and confidential computing for environment integrity; dynamic network identity and microsegmentation to limit east-west agent traffic; and a centralized control plane for permission, token, and tool governance.
  • Outlook: Product integrations and partner case studies from Nutanix, Cisco, and Intel cited in the article will be the next concrete artifacts to watch as enterprises move from pilots to broader agent deployments and require demonstrable hardware-rooted trust and control-plane telemetry.

Sources: venturebeat.com · venturebeat.com

LAION releases Big Video Dataset: 80M videos, 10M hours, 55M clips and 300M images for research use

  • What happened: LAION published the Big Video Dataset (BVD) drawn from 1.3 billion video URLs; the team downloaded 80 million videos totaling 10 million hours, extracted 55 million clips with auto-generated video and audio descriptions, and produced 300 million still images. LAION is releasing the dataset and code for research-only use and cites a paper claiming models trained on BVD outperform comparable InternVid-trained models by up to 2.1 percentage points on common video-to-text benchmarks.
  • Why it matters: The scale and multimodal alignment (video, audio, text) give engineers and researchers a large public corpus to train or finetune video-to-text and multimodal models without commercial licensing from a single vendor, enabling replication of training recipes and benchmarks at larger scale than previously available.
  • Outlook: Benchmark and model checkpoints trained on BVD, and independent comparisons to InternVid reported in preprints or conference submissions, will be the first public verification of LAION’s claimed up-to-2.1-point advantage.

Sources: the-decoder.com

Anthropic publishes a Model Hardware Standard research preview to unify device drivers for agents

  • What happened: Anthropic announced the Model Hardware Standard (MHS) research preview, a spec and driver approach that makes devices discoverable in a common format so agents can read and control lab and factory hardware. Anthropic says partners (HHMI Janelia, Genentech, Carnegie Mellon, QuEra) cut integration time from weeks or months to hours or minutes in early tests and that an open-source release is planned later.
  • Why it matters: MHS provides a single programmatic interface for heterogeneous devices, allowing agents to reuse drivers instead of rebuilding per-integration and enabling orchestration across multiple devices; that reduces integration engineering work and changes the operational surface engineers must test when moving agents into physical environments.
  • Outlook: The research preview now and the planned open-source release later are the next milestones; partner outputs from Genentech, Carnegie Mellon, and QuEra presented in Anthropic’s tests will be the immediate artifacts to inspect for real-world integration time and safety trade-offs.

Sources: the-decoder.com

OpenAI testing a Persistent mode for Codex that lets agents continue work across sessions

  • What happened: Wired reported OpenAI added Persistent mode code to Codex’s command-line tool, which the company confirmed is under test and not broadly rolled out. Persistent mode appears in Codex’s reasoning-effort menu and instructs the agent to continue working until put to sleep, create proactive follow-up tasks, use past user interactions and knowledge of the user across sessions, and message users sparingly. OpenAI says there are no immediate launch plans.
  • Why it matters: Persistent mode changes the runtime model for agents from short-lived task execution to long-running, stateful workflows that can spawn follow-up actions and consult historical interactions; that raises new engineering questions about token budgets, proactivity limits, sandbox integrity, and alignment safeguards because the agent can operate across sessions and create tasks for itself.
  • Outlook: Commits and further changes in the Codex command-line repo, plus any signals that Persistent mode is promoted into Codex desktop, ChatGPT Work, or other product channels, will be the next public indications of broader rollout and operational constraints.

Sources: wired.com