Diagrid’s Catalyst 2.0 layers a durable workflow on top of popular agent frameworks, letting a failed agent pick up exactly where it stopped without re‑running earlier steps. The feature, built on Dapr, also records each model and tool call, giving enterprises tamper‑evident logs for high‑stakes workloads.
Uber engineers rewrote Go stack handling to pre‑allocate larger goroutine stacks, eliminating costly runtime expansions. The change trimmed CPU consumption by 10% on a service that runs on over 2 million cores, translating to multi‑million‑dollar savings at scale. The technique is now being rolled out across Uber’s Go‑based services.
Netflix migrated its production LLM serving from TensorRT‑LLM to vLLM, wrapping it in NVIDIA Triton for model loading, batching and GPU scheduling. The new stack lets small models run on‑CPU while large ones use a shared Model Scoring Service, cutting latency and giving engineers full control over rollout and observability.
Traditional API gateways assume deterministic services and static schemas, assumptions that break with agentic AI. The article proposes an evolutionary architecture pattern that version‑controls contracts, adds automated fitness tests, and lets platform teams iterate safely as AI models and pipelines change rapidly. It shields services from brittle breakage and keeps delivery velocity high.
GitHub and npm introduced two key defenses: npm now places high‑impact accounts in read‑only mode for 72 hours after an email change or 2FA recovery, buying time to stop credential theft; GitHub Actions’ checkout action now defaults to blocking untrusted code from fork‑origin pull‑request triggers. The tweaks cut off early‑stage compromise vectors in supply‑chain attacks.
AWS’s GuardDuty Investigation Agent, now in public preview, auto‑correlates findings, 90‑day logs and resource topologies into risk‑rated, MITRE‑mapped reports. Accessible via console, CLI or API, it shrinks investigation time from hours to minutes, letting security teams focus on remediation instead of manual digging.
Elastic shows that fragmented SOC tool stacks cost analysts time and money, letting AI‑powered attackers finish a compromise in under a minute. It argues that only a unified, agentic platform that provides real‑time, cross‑telemetry access can let security teams use AI effectively and keep pace in 2026.
You’re caught up.
That was all 7 stories in the 28 July issue.