AI agents move from filing issues to fixing them

Agents that close the loop

The week's most consequential change came from Sentry, where Seer learned to iterate on the pull requests it opens. Previously, any change to a Seer PR meant re-running Autofix and getting a brand-new PR; now feedback from the Sentry Autofix panel, GitHub comments starting with @sentry, and review comments from humans or bots all land as follow-up commits on the existing PR, keeping the diff, thread, and history in one place. Seer also investigates failing CI checks and pushes fixes, with automatic attempts capped, so a red build can be diagnosed without a human re-prompting the agent. PR iteration is GitHub-only for now, but it reframes the agent as a participant in review rather than a one-shot code generator.

Dash0 pushed in a similar direction with automations that trigger on GitHub Actions workflow events — workflow runs and jobs, manual dispatches, check runs, and check suites, scoped by org, repo, branch, or workflow name and filtered by conclusion or status, with the automation able to investigate a failed run and push a fix that re-triggers the workflow. Datadog, meanwhile, took the opposite tack on where agents live: Bits Chat is now available in Slack, letting responders run an investigation, pull telemetry findings, page engineers, generate a fix PR via Bits Code, and close out with a postmortem notebook without leaving the channel.

SLOs and the cost of everything else

Dash0's headline feature is Service Level Objectives with OpenSLO as the serialization format. Point an SLO at any signal you already send to get remaining error budget and burn rate, draft definitions with Agent0, and convert one into an alert in a click — with the option to keep definitions in a repository and apply them via CLI or Terraform rather than clicking them together in the UI.

Cost and security got attention on the Datadog blog. Cloud Cost Management now ingests Cursor usage, breaking AI coding spend down by user and model instead of leaving FinOps teams with a single opaque number. IaC Security added custom Rego rules that run alongside Datadog's defaults across Ansible, CloudFormation, Dockerfile, Kubernetes, Terraform, and GitHub Actions, with an AI rule creator that drafts policies from natural language and an editor for testing against sample files. For teams piping security logs into a SIEM, Observability Pipelines shipped Microsoft Sentinel Packs that map Palo Alto, Fortinet, Cisco ASA, Meraki, and ExtraHop logs into Sentinel table schemas in-flight — so raw logs can be archived at full fidelity while only high-value events hit per-GB ingest.

Tracing gets agentless and OpenTelemetry-aware

Datadog's dd-trace-py had a busy week, headlined by DD_AGENTLESS_ENABLED, a single switch that sends telemetry, traces, Remote Configuration, Dynamic Instrumentation, crash reports, Test Optimization, and LLM Observability data straight to Datadog intake instead of through a local Agent. The same work brought consistent probability sampling for mixed Datadog-OpenTelemetry environments, plus a discard field on DD_TRACE_SAMPLING_RULES to fully drop rejected trace chunks, CPython GC and stop-the-world pause metrics, and a heap profiler Mem domain on by default for Python 3.12+. The library also cleaned up after itself: gen_ai attributes no longer leak onto APM spans when LLM Observability is off, and vLLM 0.14.0 tracing and google-adk startup failures were fixed, including a ModuleNotFoundError that had silently disabled all vLLM tracing and metrics. The Agent itself got a patch for a pre-flight bug that sent metrics and the API key to datadoghq.com instead of the configured site.

Elsewhere, Sentry JavaScript 10.75.0 registers an Effect v4 ErrorReporter so failures flowing through Effect's HTTP and RPC boundaries are captured automatically, with an 11.0.0 release candidate following behind it. Sentry Python fixed a Bottle serialization loop and made functions_to_trace work with streaming spans, and RUM's browser SDK now defaults service to the applicationId and captures wasm build IDs on error events. Dash0's operator patched a collector crash tied to disableReplicasetInformer, and Sentry CLI's snapshots were mostly churn aside from uploads going to the server-chosen objectstore.