Agents Stay On. Compute Gets Specialized. Robots Keep Learning.

What happened
Snowflake introduced new CoCo and CoWork Bots that operate proactively in the background, handle recurring work, interact with tools including Slack and Excel, and bring humans back into the loop when a decision or sign-off is required. The agents can also reuse shared skills, memory, and artifacts such as reports and dashboards.

Why it matters
Enterprise agents are moving beyond responding to prompts toward continuously executing work. That changes the operating model from employees occasionally using AI to organizations managing autonomous workers that retain context and act while people are doing something else.

What’s next
Snowflake is pairing this autonomous layer with governed business definitions and centralized controls through Business Ontology and Cortex AI Gateway. Enterprises will need to determine which recurring tasks can safely run unattended and which still require explicit human checkpoints.

What happened
CrowdStrike introduced an AI Kill Switch in Falcon Guardian that can terminate unauthorized agents at the gateway, prevent them from running again, and cut off enterprise network access. Guardian also adds agent-level hooks, AI traffic inspection, and controls designed to connect what an agent was told to do with what it ultimately executes.

Why it matters
The security question is shifting from detecting whether an AI system is malicious to stopping an authorized agent once its behavior moves outside the intended boundary. That requires controls beneath the conversational layer, where organizations can interrupt execution rather than simply warn the model.

What’s next
The Kill Switch is available today at the gateway, while endpoint termination is still coming. CrowdStrike is also adding observability through its planned Fiddler acquisition and automated AI security testing through its planned acquisition of Dreadnode.

What happened
PagerDuty launched Paige, an SRE agent designed to detect, investigate, and help remediate operational incidents across the new PD Reliability Platform. Paige can work with alerts, logs, deployment information, and service dependencies while exposing its reasoning and proposing remediation steps for human review.

Why it matters
Incident response is a strong fit for agentic AI because failures require rapid investigation across multiple systems, but mistakes can also create outages. PagerDuty is positioning the agent as part of a blended operations team rather than an autonomous replacement for the engineer.

What’s next
Paige and the PD Reliability Platform are generally available today, while SRE Workbench is available through PD Labs. PagerDuty is testing features that let Paige investigate multiple root-cause hypotheses, map dependencies, and preserve lessons from resolved incidents.

What happened
SambaNova and Supermicro announced plans for inference systems that combine NVIDIA GPUs with SambaNova SN50 RDUs, with an initial plan covering 10,000 SN50 chips. The architecture assigns GPUs to the compute-heavy prefill stage while SN50 hardware handles the memory-intensive decode stage that generates model responses.

Why it matters
Agentic workloads can trigger thousands of model calls, making inference speed, power, and cost increasingly important infrastructure constraints. Rather than asking one accelerator to perform every stage, the system splits the workload across hardware optimized for different jobs.

What’s next
Supermicro will integrate, validate, deliver, and support the systems, starting with NeoCloud and sovereign-AI customers before expanding to enterprise data centers. The companies are targeting broader inference deployments during 2027.

What happened
OpenAI demonstrated Astra helping design a physical device across mechanical design, PCB layout, component selection, firmware, and software integration, with two team members producing the prototype design in under a week. OpenAI also says its hardware teams are increasingly using Codex and ChatGPT for CAD, manufacturing drawings, simulations, and engineering reviews.

Why it matters
AI coding agents are beginning to cross the boundary from software into products that must obey geometry, materials, tolerances, electronics, and manufacturing constraints. That could make engineering iteration significantly faster, but physical mistakes are harder to undo than incorrect text or code.

What’s next
OpenAI acknowledged that Astra still missed production limits, tolerances, and assembly details that required multiple human review cycles. The company is expanding industry collaborations, including new research with Mazda on vehicle aerodynamics and crash-safety engineering.

What happened
AGIBOT unveiled GO-3 Preview, an embodied foundation model designed to learn across pre-training, post-training, and real-world deployment. The company says the model ranks first on RoboDojo, RoboTwin, and RoboColiseum and combines task understanding, prediction of future states, and physical action generation.

Why it matters
Robot intelligence cannot stop improving when training ends. Physical environments vary constantly, so systems that can incorporate experience from deployed robots could create a feedback loop where real-world work continuously improves future robot behavior.

What’s next
AGIBOT plans to fully open-source GO-3 Preview by the end of 2026. Independent developers will then be able to test whether its reported benchmark leadership translates into reliable manipulation and generalization outside AGIBOT’s own deployments.

What happened
NVIDIA outlined an “AI-defined fab” architecture combining accelerated simulation, physics-informed AI, digital twins, and autonomous agents across semiconductor manufacturing. Examples include Siemens using cuLitho to accelerate key lithography simulations by up to 17x, Lam Research using PhysicsNeMo to predict plasma behavior in milliseconds instead of days, and Advantest using agents to reduce model-selection workflows from hours to minutes.

Why it matters
Physical AI does not have to look like a humanoid robot. Semiconductor factories are complex physical systems where AI can simulate processes, predict material behavior, automate engineering steps, and test operational changes before they reach expensive production equipment.

What’s next
NVIDIA is pushing toward a connected loop where production data improves models, models evaluate proposed changes, and engineers feed validated decisions back into the fab. The value will depend on whether faster simulations and automated analysis ultimately improve yield, throughput, or development cycles.

AI is becoming less like software people occasionally consult and more like infrastructure that continuously executes work. That raises the importance of kill switches, human approvals, specialized inference systems, and feedback loops that improve agents after deployment. The next advantage will come from making autonomy persistent without making it uncontrollable.

See Where Robot Intelligence Breaks

AGIBOT cites the public RoboDojo leaderboard as one of the benchmarks used to evaluate GO-3 Preview.

Try this: Open the leaderboard and switch between Generalization, Precision, Long-Horizon, and Memory. Compare several of the leading robot models instead of looking only at the overall score.

Does the same model lead every category? Pay attention to how dramatically rankings can change when the task shifts from precise manipulation to remembering information or completing longer sequences.

Insight: There may not be one “best” robot brain. Physical AI performance depends on the type of work, making task-level evaluation increasingly important before deploying robots into production.