Agent tracing goes GA as AI observability matures

Agent observability grows up

The headline of the week is Sentry's Agent Tracing reaching GA. It captures model calls, tool executions, sub-agent handoffs, token usage, and cost as spans inside your existing Sentry traces, and the GA adds conversation IDs that expose the full message exchange between agents and users. The overview now shows readable conversation titles alongside charts of cost, conversation count, and total messages over time, and the timeline and transcript views render tool calls inline with input previews and durations. For teams debugging why an agent went sideways, this is the difference between guessing and following the actual run.

Dash0 spent the week pushing its Agent0 assistant in a similar direction. Its AI Coding Insights became AI SDLC Insights, pulling cost, adoption, productivity, and tool metrics onto a single page with an overview of total spend, active users, cycle time, average cost per user, and the share of assisted pull requests, plus a pipeline view tracking where merged PRs slow down between coding, review, and merge. Agent0 also gained per-thread model levels and reasoning effort — four tiers from Light to Premium with credits charged on what a turn actually consumed — and custom skills that let admins encode team runbooks and invoke them with a slash command.

Governed automations and GitOps

Dash0 rounded out the assistant work with a permissions change: anyone with read access to a dataset can now create, edit, and delete Agent0 automations in it, mirroring how dashboards already worked, while automations managed through the API or Terraform stay read-only in the UI. For teams doing GitOps, a new sync-assets GitHub Action keeps dashboards and check rules in step with a repo, so assets deleted from source also disappear from Dash0 on every push, manual run, or schedule. On the runtime side, the Dash0 Lambda extension now extracts HTTP semantic-convention attributes itself, reading raw invocation and response payloads so handler spans get http.request.method, url.path, http.route, and status codes across every runtime. The operator, meanwhile, added OpenShift compatibility alongside TSA CRD support, node UID self-monitoring telemetry, and fixes for a memory leak and Signal Control counters, with the changelog also covering the Helm install path for enabling SecurityContextConstraints.

Datadog's week: Kafka, experiments, and sampling

Datadog launched Kafka Console for end-to-end Kafka troubleshooting, correlating configuration and schema changes with throughput and lag graphs and letting authorized users retrieve messages by partition and offset, with Bits AI able to build a hypothesis tree and recommend fixes. The company also detailed how it rebuilt Eppo into Datadog Experiments, extending CUPED variance reduction to segments and percentile metrics and redefining global lift as coverage times local lift, and separately described analyzing experiments inside ChatGPT via a plugin. A NetFlow duplication explainer and a product launches workflow rounded out the blog.

On the SDK side, Datadog's JavaScript tracer added Turbopack instrumentation and Azure Durable Functions support on the current line, with the same instrumentation landing on the older line as well. Both lines also picked up trace discarding via DD_TRACE_SAMPLING_RULES and proxy environment variable support for tracing and OTLP export. The Python tracer fixed a pytest-xdist slowdown on shallow checkouts, and the Agent fixed fast network path test billing and a containerd SBOM snapshot leak while releasing the leader-election lock cleanly on shutdown.

Route attributes and tracing polish elsewhere

Sentry's Python SDK added http.route attributes to Django and to Bottle, Falcon, and aiohttp, plus ClickHouse Driver streaming trace breadcrumbs. Its JavaScript SDK folded Gemini reasoning tokens into Vercel AI usage and shipped a migration guide iteration for the upcoming major version. Axiom's Go SDK dropped legacy query support.