Sentry and Datadog push AI instrumentation, agent governance, and supply chain defense

AI instrumentation lands everywhere at once

The week's clearest thread is telemetry for AI workloads moving from manual wrapping to defaults. Sentry shipped a new mcpServerIntegration that wraps every McpServer instance by default, so teams no longer call wrapMcpServerWithSentry by hand — and the same SDK line auto-instruments Hono requests with route-named spans and turns on dedupeIntegration for Node and Bun. Sentry also gave TypeSafe Jev first-class tracing in two SDKs: the JavaScript SDK's typesafeIntegration traces @typesafe-ai/sdk calls and extends experimental_evaluate coverage through the Vercel AI Gateway, while the Python SDK's TypeSafe integration captures gen_ai input and output messages and stops dropping events when a processor raises.

Datadog's tracing libraries mirrored the push. The JavaScript tracer added versioned prompt config and message placeholders in managed prompts for LLM Observability, and made Test Optimization failure media uploads opt-out while supporting streamed Anthropic responses in AppSec. The Python tracer tagged cancelled or errored Bedrock streamed requests with partial output and estimated tokens instead of dropping them, and fixed a macOS subprocess deadlock and a Linux memory leak under OTel context propagation.

Agents that act, not just answer

Sentry's Seer Agent crossed a line: it can now write to Sentry through the API under the user's own permissions, building dashboards, creating monitors, resolving issues, and saving explore queries with per-action write-scope confirmation. The same release added Seer Connectors that pull infrastructure data from Datadog and Google Cloud via API keys — so the assistant reasons over third-party telemetry, not just Sentry's own. Dash0's Agent0 got the operational counterpart: threads keep running when you switch away, with a bottom bar listing running and finished work.

Dash0 also tightened who can change what. Automations are now editable only by admins and the members or teams they're shared with — previously anyone with dataset read access could rewrite or delete any automation. New organizations can also pin telemetry to AWS us-east-1 as a second US region, and the app got a calmer layout with fewer separator lines plus pinnable pages.

Supply chain defense goes org-wide

Datadog's biggest security story was Suppy Chain Firewall graduating to an organization-wide control inside Code Security: teams define allowlists and blocklists once and apply them across workstation and CI installs, with a new GitHub action and retroactive scanning of recorded installations against updated malicious-package intelligence — directly aimed at worms like Shai-Hulud 2.0 that harvest credentials during npm install.

Elsewhere in the stack, the Datadog Agent made its process manager the default service manager on Linux and switched log sending to fixed concurrency, while the RUM browser SDK enabled partial view updates by default for everyone and stripped query strings and fragments from Meta record URLs. Datadog's research side released Distributed DataFusion, an Apache-licensed extension of DataFusion that runs one query across many machines, and detailed an async-aware Python profiler that models parent-child asyncio tasks and cuts overhead by more than 60 percent. Two guides rounded out the week for data teams, covering Databricks jobs, clusters, and model serving monitoring and querying Databricks system tables directly.