The Engineer

Microsoft released Execution Containers on October 7, Techpresso reports. They limit an agent's file, network and interface access from outside the agent, so the agent can't widen its own access. CSO reports that AWS's new Strands Box uses OS isolation instead of a VM. CSO also reports that AWS agents have repeatedly been tricked into revealing credentials. If your agents hold their own secrets or police their own egress, they fall below the bar these vendors now describe.

LinkedIn connected a shared data context layer to its analytics agent. In one test, acceptable answers rose from 25.9% to 79.5%, per its engineering write-up. Median time-to-answer fell from 10.7 to 4.7 minutes. If your agents query a warehouse, documenting your data is likely a cheaper accuracy win than upgrading the model. LinkedIn hasn't published the grader or question set, so replicate the result on your own eval first.

Microsoft, Liquid AI and Subquadratic each shipped a decision model within weeks of each other. A decision model scores a fixed set of answers in one forward pass and generates no text. Microsoft-Decision-1 is priced at $0.042 per million input tokens with 85 ms median latency, per AI Breakfast. That lets your routers, verifiers and continue-or-stop gates become a cheap component separate from the generator. Separately, The Information reports that the top-10 models' share of OpenRouter usage fell from 55% to 14%.

Anthropic demoed dynamic workflows at AI Engineer New York, Latent.Space reports. Claude writes an orchestration program that launches subagents, replacing the hand-configured roster in Managed Agents. The model can't see its budget while it runs, so your session budget works as a circuit breaker, not a plan. Anthropic says flows cost 'much more' than single-agent runs but gave no multiple.

Bloomberg reports that electricity is now the biggest roadblock to AI data center construction, and that Oracle is trucking gas to its sites. Lumentum says its optical components are sold out through early 2029. For your capacity plan, treat a provider's online-by date as an energization date, and keep a second provider path tested. Microsoft says Maia 300 delivers 40% more output per unit of power than leading Nvidia systems, but the figure is self-reported and the chip isn't for rent.