Agent observability matures as Sentry and Datadog ship session replay and LLM debugging tools

Agent observability goes mainstream

Two of the biggest names in observability made agent observability more practical this week. Sentry opened Conversations to all users in open beta, giving developers a readable timeline of past agent sessions — messages and tool calls together, not scattered across separate logs. The feature fills a gap that anyone building on LLM frameworks has felt: understanding why an agent did what it did without stitching together traces and chat logs by hand.

Meanwhile, Datadog published a guide comparing guardrail placement strategies across Amazon Bedrock Agents and self-orchestrated agents using AI Guard — timely given the indirect prompt injection demo used in the post. A separate guide demonstrated using LLM Observability Experiments to track an AI agent's autonomous iteration cycle: one agent ran 23 experiments to improve a SQL query optimizer from 54% to 86% accuracy, tracking every hypothesis and failure along the way.

Datadog tightens Bedrock and Langchain instrumentation

Several Datadog APM releases addressed gaps in LLM framework coverage. Python APM v4.8.5 fixed Bedrock Converse guardContent blocks that had been dropping user input from traces when text was wrapped in guard responses, and deprecated DD_TRACE_INFERRED_SPANS_ENABLED in favor of DD_TRACE_INFERRED_PROXY_SERVICES_ENABLED. Python APM v4.8.6 fixed Langchain Bedrock inference profile attribution — spans were being tagged with the profile's ARN instead of the actual LLM model. On the JavaScript side, APM v5.104.0 fixed an unfinished CONNECT span in undici instrumentation and resolved a delta temporality issue with OTLP counter exports.

The Datadog Agent v7.79.0 upgraded JMXFetch to 0.52.0, adding mappings for Generational Shenandoah GC, but macOS users need to note a breaking change: the Agent now installs as a system-wide LaunchDaemon under a dedicated _dd-agent user instead of a per-user LaunchAgent, and existing installations must reinstall.

Sentry tightens security and Dash0 improves database debugging

Sentry CLI 3.4.3 disabled Xcode Info.plist preprocessing by default across releases propose-version, send-event, and react-native xcode commands — a security hardening move to prevent project-controlled compiler settings from reaching cc during release auto-discovery. The same change shipped on the 2.x line as version 2.58.6. Sentry also fixed a data scrubbing regression that prevented http.query, url.query, and url.full span attributes from being scrubbed.

Dash0 added a dedicated database query widget to the span sidebar, syntax-highlighted and with prepared statement parameters filled in, so you no longer have to hunt through the attributes tab for slow query debugging. The Dash0 Operator 0.141.0 added a captureSqlQueryParameters flag for finer-grained control.

Smaller releases worth knowing about

Datadog published several guides covering practical ops workflows: budget forecasting for Cloud Cost Management, a monitor audit framework for reducing alert fatigue, natural language queries for exploring metrics in plain English, explain plan correlation for PostgreSQL in Database Monitoring, and a guide to reducing CVE noise with OpenVEX assessments.