Opus 5.5 adds conversation-replay checks
New model documentation makes retained instructions and thinking-block compatibility part of the compaction migration review.
Original report: Claude adds on-demand compaction for long-running agentsAIOps / Observability / Infrastructure
News, with an operator’s perspective.
Developments in AI and the infrastructure behind it. What changed, how it works, and what it means for the people running it.
Our editorial standardsFollowing the story
Developments since our original reporting, with dated updates and supporting sources.
Releases, research & operational impact
Latest coverage
AI infrastructureNews analysis
NVIDIA tests GPU cluster readiness with real workloads
NVIDIA introduces workload-driven GPU cluster validation. Choose the acceptance claim and thresholds before interpreting a completed run as readiness.
Read the analysis
Kubernetes operationsNews analysis
NVIDIA introduces NodeWright for Kubernetes host updates
NodeWright brings Kubernetes-aware host changes into a declarative workflow. Check node selection, interruption limits and recovery before a fleet rollout.
AI modelsNews analysis
Claude Opus 5.5 lowers prices and changes the migration check
Anthropic releases Opus 5.5 with lower token prices. Evaluate operational exceptions and integration behavior before changing an agent workflow.
Inference infrastructureNews analysis
TensorRT’s multi-GPU serving results put capacity in focus
NVIDIA’s September 21 demonstration accelerates one video-generation workload across eight GPUs. A service decision still needs the same offered load and completed-work count.
Agent evaluationNews analysis
SWE-Serve tests coding agents through live inference
NVIDIA’s new SGLang benchmark finds patches that pass local checks but fail a running server. Its result is a reason to examine what a green test actually covers.
Local AI agentsNews analysis
Antigravity brings local models to its agent SDK
Google adds local Gemma and LiteRT workflows. Running the model on a workstation still leaves a separate decision about what its tools may read, change or send.
AI inferenceNews analysis
NVIDIA’s confidential inference results need a workload boundary
NVIDIA reports low confidential-computing overhead on an eight-GPU test. Separate that measured performance result from your own security and recovery claims.
AI infrastructureNews analysis
NVIDIA Topograph puts GPU placement evidence in focus
New Topograph guidance connects physical network topology to workload schedulers. The operational check is whether the scheduler’s view survives cluster change.
AI model operationsNews analysis
GPT-6 Sol and Luna change the agent cost calculation
Lower model prices and new caching controls make agent pilots cheaper to run. Measure accepted work before expanding the workload.
AI infrastructureNews analysis
New AI energy alliance puts workload flexibility on the SRE agenda
Google, NVIDIA and Emerald AI are backing flexible data centers. Operators will need to define which AI work can yield power, and how it recovers.
Agent infrastructureNews analysis
Claude adds on-demand compaction for long-running agents
The Messages API can now summarize conversation history when an application chooses. SRE teams should test which operational constraints survive.
AI agent governanceNews analysis
Claude’s Chrome transcripts enter the enterprise audit trail
Anthropic has added Chrome-session transcripts to its Compliance API beta. Responders gain another evidence source, with important gaps to understand.
AI operationsNews analysis
Splunk agent observability: follow the work, then the bill
Splunk’s September release connects agent evaluation and cost monitoring. The useful test is whether a responder can trace spending back to an outcome.
AI observabilityNews analysis
OpenTelemetry’s Kubernetes milestone matters for AI agent tracing
The stable Kubernetes attributes processor helps connect AI-agent traces to their workloads. Check the metadata joins before upgrading.
AI infrastructureNews analysis
Kubernetes Memory QoS: what AI inference teams should test
Memory QoS reaches beta in Kubernetes 1.37. For AI inference teams, the rollout question is how host-memory policy affects serving latency and neighboring workers.