Sentry takes agent tracing GA as Datadog reshapes tracing defaults

Agent observability grows up

The week's headline is Sentry's Agent Tracing reaching GA. For anyone running LLM agents in production, the pitch is that agent runs are no longer a black box bolted onto your telemetry: model calls, tool executions, sub-agent handoffs, token usage, and cost all land as spans inside the traces you already collect. Conversation IDs tie a request to the full message exchange between agent and user, and the overview now renders readable conversation titles alongside charts of cost, conversation count, and total messages over time. Tool calls appear inline in the timeline with input previews and durations. The framing is failure-first — you start from a bad response and walk back through the trace — which is the right instinct. Cost and latency dashboards for agents are commodity now; connecting a degraded answer to the specific tool call that caused it is not.

On the SDK side, Sentry's JavaScript 10.74.0 folded Gemini reasoning tokens into Vercel AI usage counts and set user_agent.original on all browser spans, tightening the filtering story for agent traffic that flows through Vercel's AI SDK. The 11.0.0-beta.2 line also published another iteration of its migration guide — still explicitly work in progress, so plan accordingly if you're tracking the major.

Datadog's tracing defaults get sharper

Datadog spent the week making JS tracing more configurable. Two releases — dd-trace-js 5.127.0 and 6.16.0 — shipped in parallel across the 5.x and 6.x lines, both adding Turbopack instrumentation for Next.js server bundles, Azure Durable Functions orchestration triggers, and ai-vercel streaming support for AI Guard, plus an OpenTelemetry meter provider shutdown. A separate APM release brought trace discarding via DD_TRACE_SAMPLING_RULES, with a matching 5.126.0 entry. Discarding traces at the scope level — rather than relying only on sampling at the head — gives teams a way to shed noise from chatty routes without losing the trace that matters. Both also honor proxy environment variables now, which matters for anyone tracing through a corporate egress.

Smaller but real: dd-trace-py v4.11.8 fixed a nasty CI Visibility case where pytest-xdist sessions on shallow Git checkouts spent excessive time unshallowing the repo, occasionally timing out entirely. And the Agent 7.83.1 release cleaned up two billing-and-resource bugs: fast network path tests were being billed to customers, and containerd SBOM scans leaked their view snapshot and lease when cancelled or timed out.

Routing, runbooks, and the platforms around them

Sentry's Python SDK spent the week adding http.route attributes across frameworks — 2.69.1 brought Django on board and stopped aiohttp from dropping SigV4-signed propagation headers, while a companion release extended the same attribute to Bottle and Falcon and added breadcrumbs to the ClickHouse driver's streaming trace lifecycle. Route-level attribution is unglamorous but it's what makes grouping and alerting usable.

Elsewhere, Dash0 let teams write custom skills for Agent0, teaching the assistant your runbooks, naming conventions, and triage order — the missing piece for teams whose institutional knowledge doesn't live in a docs site. Its Operator 0.155.0 added OpenShift compatibility and fixed a memory leak in DelegatingZapCore. Axiom's Go SDK v0.37.0 removed legacy query support, a breaking change worth checking before you bump. Datadog, meanwhile, launched Kafka Console for correlating config and schema changes against throughput and lag, and continued its Eppo-to-Experiments rebuild writeup.