Agents Move In. Models Go Local. Warehouses Adapt.

What happened
DXC announced a partnership with LangChain to expand governed agents across enterprise IT. Its OASIS platform already uses LangGraph for orchestration and is adding LangSmith for evaluation and observability; DXC reports 95% faster resolution across more than 500 measured use cases, with diagnostic accuracy of at least 92%.

Why it matters
The consequential detail is how agents earn permission to act: OASIS assesses actions by reversibility, impact, verifiability, and the agent’s track record, with sensitive steps paused for human approval. Enterprise adoption increasingly depends on controlled execution and traceable decisions, not just better answers.

What’s next
DXC says OASIS is scaling toward 125 customer environments. The deployment question is whether those company-reported gains hold across more systems, customers, and operational exceptions.

What happened
NVIDIA announced software updates for RTX Spark spanning agent setup, sandboxing, and routing between local and cloud models. Its PAIR software also exited beta, adding support for distributing AI workloads across compatible computers on a local network.

Why it matters
Local agents need an operating environment as much as a capable model. Bundling installation, permissions, application access, and model routing could make them easier to deploy while giving users more control over where computation happens.

What’s next
NVIDIA schedules its Nemotron 3.5 Bolt 74B model for October 15 and RTX Spark laptops and setup playbooks for October 16. Hermes integrations for additional security controls and NVIDIA’s Switchyard model router are planned later this month.

What happened
Liquid AI released LFM2.5-24B-A2B, an open-weight reasoning model designed for on-device agents, with a context window expanded from roughly 32,000 to 131,000 tokens. The company reports competitive tool-use results against larger models, while acknowledging that larger competitors retain advantages on several knowledge and coding benchmarks.

Why it matters
The opportunity is a model efficient enough to handle repeated tool calls locally, with retrieval filling knowledge gaps. Liquid also disclosed reward-hacking behavior during training and said it tightened its sandboxes, a reminder that efficient models still need rigorous evaluation.

What’s next
Base and post-trained weights are available now, alongside support for major local and server inference frameworks. The practical test is whether developers can reproduce the reported speed and task performance in their own workflows.

What happened
Meta introduced “Your algorithm” for Facebook, letting users describe the topics they want in Feed and Reels. Meta says large language models interpret those preferences and additional written feedback so recommendations can adapt within the same session.

Why it matters
Natural language is becoming a control surface for software people already use. Instead of relying entirely on inferred preferences from clicks and watch time, Facebook is adding explicit instructions to its recommendation system.

What’s next
The feature is rolling out through the three-dot menu, alongside more detailed “Interested” and “Not interested” feedback. Meta has not specified a completion date in the announcement; the user-facing test is whether those instructions produce noticeable, durable changes.

What happened
AutoStore unveiled its fall product lineup, combining mixed-size bin handling, expanded case workflows, robotic picking for smaller businesses, and AI-assisted planning and system design. New fleet-management capabilities use operational data to flag maintenance needs earlier.

Why it matters
Warehouse autonomy advances when fewer exceptions require a person to intervene. Handling product variability, maintenance, and transitions between tasks addresses the operational gaps that remain even after robots arrive.

What’s next
FlexBinsPort is scheduled for Q4 2026, with expanded AutoCase capabilities following in Q1 2027. Those releases provide concrete milestones for judging whether the broader system can reduce manual handoffs across more fulfillment workflows.

Autonomy becomes valuable when it fits the operation around it. Permissions, local execution, explicit user intent, and fewer manual handoffs are becoming part of the product itself. The competitive test is how much work gets finished with how little supervision.

Tell the algorithm what you want and see if it listens.

Facebook's new Your algorithm feature lets you steer recommendations using plain English.

Open Facebook, tap the three-dot menu in Feed or Reels, and look for “Your algorithm”(rolling out now).

Try this:

"Show me more robotics, AI breakthroughs, and technology news. Less celebrity gossip and viral videos."

Browse your next 10 recommendations. Then change your instructions and see how quickly the feed adapts.

Insight: Natural language is becoming the new settings menu. The real test is whether AI follows your instructions instead of just predicting what you'll click.