Conversational observability and agentic alert response

Conversational observability reaches GA

The biggest story this week is observability that answers questions in natural language. Datadog Bits Chat is now generally available — a conversational interface that lets you search, visualize, and take action across Datadog using plain English, with access to metrics, logs, traces, dashboards, monitors, and documentation. It generates new dashboards and notebooks on the fly, correlates signals across services to surface likely root causes, and answers best-practices questions by pulling from Datadog's own docs. It works in the web app, mobile, and Slack.

Complementing the query side, Bits Agent Builder lets teams create custom AI agents that can reason through alerts and take action automatically. These agents have access to 2,000+ prebuilt actions and can be chained together or triggered conditionally — no hardcoded logic branches required. For AI agents that need programmatic access, Pup CLI ships as a single binary with 200+ commands across 33 product domains, using OAuth-scoped tokens instead of long-lived API keys.

Production-grounding AI optimizations

Datadog published DODO (Datadog Observability-Driven Optimizer), a technique for grounding LLM-driven code optimization benchmarks in live production telemetry. By matching benchmark execution shape to production CPU profiles with ≥98% similarity, DODO ensures that AI agents optimize the right code paths. Applied to one critical service, three optimizations reduced CPU cost by over 8%, saving on the order of 10,000 cores continuously. The post is a strong read for anyone building or evaluating AI coding agents.

Security monitoring and supply chain defense

For enterprise security teams, Datadog's Claude Compliance API integration brings Claude Enterprise activity data into Cloud SIEM with prebuilt detection rules, OCSF-aligned log pipelines, and compliance dashboards covering sign-ins, admin actions, API key lifecycle, and configuration changes. Separately, a Datadog Security Labs post details the landscape of supply chain attacks targeting widely used dependencies like axios and LiteLLM, with guidance on detection at scale.

SDK and infrastructure hardening

Across the SDKs, memory leak fixes dominated. The Python APM tracer v4.10.1 and v4.8.8 both fix unbounded native call-site metadata growth and PeriodicThread reference cycle leaks; v4.8.8 additionally fixes an SCA memory accumulation issue. v4.10.2 corrects Claude Agent error attribution (no more "unknown" error categories) and a race condition dropping service names under uWSGI. The JS tracer v5.106.0 adds Bedrock Converse/ConverseStream support and fixes OTLP spans to include full 128-bit trace IDs. The Datadog Agent 7.79.2 patches CVE-2026-46680 in containerd and fixes DockerUtil initialization failures from malformed CIDR responses. Sentry's JavaScript SDK 10.55.0 stabilizes the dedicated @sentry/hono SDK; 10.56.0 adds Redis integration for Deno. Sentry's CLI 3.5.0 adds local snapshot comparison and download commands.

Behind the scenes: PostgreSQL on Kubernetes

Datadog's engineering blog detailed a gameday simulation that exposed a PostgreSQL HA weakness on Kubernetes: when network latency spiked in one availability zone, replication lag grew unchecked and failover became unsafe. The rearchitecture uses synchronous replication for failover candidates coordinated by Patroni, with benchmarked trade-offs between durability and latency — useful reading for anyone running stateful workloads on K8s.