November 2026 Summaries

Datadog engineers investigated why internal AI applications were not consistently benefiting from LLM prompt caching despite having it enabled, using Cloud Cost Management to identify provider spending and Agent Observability to trace individual model requests, token usage, latency, and cache reads and writes. They found that caching requires requests to share an identical beginning, so unstable details such as timestamps placed early in system prompts, nondeterministic ordering of tools or files from Go maps, missing provider-specific cache controls, and use of unsupported API compatibility routes could all prevent reuse. The team recommended placing stable instructions before changing context, sorting generated tool and file lists, verifying that cache settings reach provider request bodies, and using supported APIs where appropriate. They also stressed that improvements should be evaluated across complete tasks rather than isolated calls, accounting for cache-write fees, output and retry costs, response time, task accuracy, and provider billing data to confirm actual savings.

Datadog developed FETCh, a system that translates TLA+ formal specifications into Rust runtime monitors, to continuously check whether production behavior remains consistent with the models used to verify distributed systems during design. The effort arose from Husky, Datadog’s event database, where a routing specification had gradually diverged from implementation changes: a placement-logic exception allowed overlapping placements to persist, creating a latent risk that the backfill system could fail after placement-data loss. Rather than monitoring only high-level invariants, FETCh validates individual state transitions against specifications, using TLA+ refinement to connect observable production models with bounded, model-checkable versions. To operate at scale, Datadog split Husky validation into local per-process routing monitors and a global placement-storage monitor, using sidecars and RocksDB-backed state structures. The deployment found the long-running placement conflict despite no observed normal-operation impact, while adding roughly 2% CPU and under 0.2% memory overhead at the 90th percentile. FETCh currently supports only part of TLA+ and requires instrumentation and system-specific handling of global state, but Datadog presents continuous specification validation as a way to detect model drift and preserve the value of formal methods as production systems evolve.

Oracle AI World 2026 highlighted the integration of AI agents into established business systems, with Oracle announcing a collaboration with OpenAI that connects ChatGPT, Codex, and OpenAI models to Oracle applications, data, and cloud infrastructure through Oracle AI Workspace. Oracle also introduced Fusion Companion and Fusion Agentic Home to enable users to delegate Oracle Fusion business workflows to AI agents, increasing the importance of monitoring underlying jobs, integrations, and user activity. Datadog presented its observability capabilities for Oracle Fusion Cloud Applications, Oracle Database, and OCI, including monitoring for Enterprise Scheduler Service jobs, audit logs, synthetic user testing, database performance, and cloud costs. Multi-cloud expansion was another key theme, as Oracle AI Database services continue to operate within AWS, Azure, and Google Cloud environments, while Peloton Consulting Group described using Datadog to monitor its Oracle Fusion-based AI workflows. Oracle additionally announced plans to pursue FedRAMP Class D certification for its multi-cloud database services supporting U.S. government workloads, with availability planned from 2027.

DatadogInstrumentation is a Kubernetes custom resource that centralizes workload-specific observability settings for integration checks, log collection, and APM, allowing application teams to manage them through familiar manifests, version control, pull requests, and GitOps workflows while platform teams retain responsibility for the Datadog Agent. Using a checkout-service example, the resource targets a Deployment and defines HTTP and OpenMetrics checks, log metadata and sensitive-data masking rules, and Java tracing settings in structured YAML rather than distributing configuration between pod annotations and Agent resources. Check and log updates can be validated by the Cluster Agent and applied at runtime without restarting the application or Agent, whereas tracing, profiling, and other instrumentation changes take effect only for newly created pods and therefore require a rollout. The feature requires specified recent versions of the Datadog Agent, Cluster Agent, and Operator or Helm chart, and can be installed through the Datadog Operator or Helm chart alongside its supporting configuration guides.

Datadog’s Translate and Import tools are designed to ease migrations from Splunk by converting SPL-based searches, alerts, and dashboards into Datadog assets without requiring manual rewrites. Translate handles individual queries, provides step-by-step syntax mappings, validates output against live logs, and offers actions to open searches or create monitors, while Import processes exported Splunk assets in bulk, converting alerts into draft monitors and dashboards into draft widgets for review before publication. A centralized inventory tracks translated, skipped, and published assets across phased migrations, with unsupported or Splunk-specific configurations flagged rather than converted unreliably. The migration process can also help teams adopt native Datadog capabilities such as Tap to Parse, subqueries and Reference Table joins, scheduled CSV reports, and Notebooks for more complex analyses. Available at no additional cost to most Datadog Log Management customers, the tools support both incremental validation and large-scale asset migration.

Datadog Online Experiments, integrated with Agent Observability and Datadog Experiments, enables teams to evaluate AI-agent changes such as new prompts, models, retrieval strategies, and tools using randomized live production traffic. While offline evaluations remain useful for testing predefined cases and preventing known regressions, online experiments reveal unanticipated user behavior and measure quality alongside operational trade-offs including latency, token use, cost, and workflow errors. Teams can use existing evaluators, confidence intervals, feature-flag traffic allocation, and execution traces to compare a control variant with a changed version and investigate the causes of observed differences. In an internal comparison of GPT-5.6 Luna and GPT-6 Luna for Datadog’s Insights agent, Luna 6 reduced estimated costs by 44.3% while producing similar quality scores, but traces showed it generated invalid tool parameters more frequently, requiring retries. Datadog addressed that issue before proceeding and identified the failures as candidates for future offline regression tests, illustrating how online findings can strengthen both production experimentation and offline evaluation coverage.

Datadog Agent Observability now integrates with Fireworks AI, enabling teams building agents with hosted open models such as GPT-OSS, Kimi, GLM, Qwen, and MiniMax to monitor application behavior, token use, traces, and quality evaluations in one platform. After connecting a Fireworks AI account and supplying an API key, users can employ available models for LLM-as-a-judge evaluations that assess production outputs for criteria including faithfulness, relevance, sentiment, hallucinations, and safety, while selecting models based on cost, latency, scale, or domain-specific fine-tuning needs. The integration also supports Datadog’s Agent Observability Playground, where developers can test prompts and real customer questions against Fireworks AI models, compare response quality, token counts, and latency, inspect full traces, and validate prompts against larger datasets before deployment. Evaluation results can be viewed alongside observability data, filtered over time, and used to identify regressions or trigger alerts, while the tested configurations can be exported into applications and monitored using the same models in production.

Datadog RUM Profiling is an add-on to Real User Monitoring that links CPU profiling data to performance events from real browser, iOS, and Android users, helping developers identify the code responsible for slow launches, unresponsive applications, hangs, ANRs, and critical workflow delays. It attaches profiles to events such as Time to Initial Display, Time to Full Display, application startup, long tasks, and developer-defined RUM operations like checkout or authentication, allowing engineers to inspect flame graphs, call stacks, and thread timelines without reproducing problems locally. Aggregated views help teams distinguish isolated incidents from recurring bottlenecks, compare slow and typical sessions or application versions, and measure improvements after optimization. To limit production overhead, mobile profiling is sampled and configurable through RUM Remote Configuration, with Android using ProfilingManager APIs and potentially system traces for ANRs, while iOS captures profiles around launches, hangs, and long tasks. Datadog presents browser and mobile profile data in a consistent interface alongside RUM data, and Bits AI plus the Datadog MCP Server can assist with identifying root causes and querying problematic sessions.

Datadog’s generally available Electron SDK extends Real User Monitoring beyond Chromium renderers into Electron’s Node.js main process, addressing observability gaps for desktop applications where authentication, synchronization, IPC, network orchestration, and OS integrations increasingly occur outside the UI layer. By combining main-process errors, crashes, HTTP requests, views, and renderer activity into a single RUM session, it enables teams to investigate issues such as background-sync failures, stale interfaces, and unexpected application exits with shared context. The SDK automatically captures unhandled exceptions and native minidump crashes, identifies the originating runtime and version details, and records Node.js network activity alongside user actions and renderer resources. It also integrates with Datadog’s Node.js tracing library to link IPC messages, command execution, and backend calls to APM traces, allowing engineers to follow work from a user interaction through the main process to backend services. Setup involves initializing the @datadog/electron-sdk in the main process and the Datadog Browser SDK in each renderer, with support available for Electron 39 and later.

Datadog Export Pipelines enable teams to continuously export Real User Monitoring and Product Analytics event data to Amazon S3, Azure Blob Storage, or Google Cloud Storage without maintaining custom scripts or scheduled jobs. Configurable through the Datadog UI, pipelines can export selected events in Parquet or JSON formats, organize files by date and hour, and provide status visibility and controls for updating, pausing, or disabling exports. The exported data can be loaded into platforms such as Snowflake, BigQuery, and Databricks for custom analytics, joined with business datasets, used in machine learning workflows for applications such as personalization or churn prediction, or retained under organization-specific compliance and lifecycle policies beyond Datadog’s standard retention periods.

Datadog Prompt Management in Agent Observability provides LLM application teams with a centralized registry for creating, versioning, deploying, and retrieving prompts independently of application releases. Prompts retain stable IDs and version histories, allowing teams to review changes, compare versions, roll back problematic updates, define runtime variables, and register prompts already captured through Prompt Tracking. Applications can retrieve appropriate prompt versions at runtime, with Python support and optional local fallbacks, while teams can deploy different versions across development, staging, and production environments or selectively target traffic using Datadog Feature Flags. When used with Prompt Tracking and Agent Observability, the service links prompt versions to LLM spans, enabling analysis of call volume, latency, token use, cost, and traces to assess how prompt updates affect application behavior.

Datadog has introduced AI Guard, a generally available runtime security product designed to protect AI agents from threats such as indirect prompt injection, sensitive-data exposure, and unintended tool use. Unlike traditional permission-based controls, AI Guard operates inline within the agent decision loop, using session history, custom policies, application traces, and infrastructure context to assess an agent’s intent at each step and block or sanitize unsafe prompts, outputs, tool calls, and results before execution. It also discovers agents and their connected models, datastores, and services, masks PII and secrets in real time, and creates investigation signals that reconstruct the events leading to blocked actions. By integrating security findings with Datadog traces, logs, telemetry, remediation workflows, Jira, and SIEM systems, the platform aims to help engineering and security teams investigate incidents, enforce policies in existing development workflows, and scale agent deployments without proportionally increasing risk.

Datadog RUM Managed Archive extends the RUM without Limits model by storing all ingested user sessions, including those excluded by retention filters, in lower-cost managed storage for one, three, or six months. Archived sessions retain searchable metadata such as timestamps, session IDs, user IDs, and account IDs, enabling teams to locate relevant historical activity without fully indexing every session. When needed for support, debugging, incident retrospectives, compliance, or audits, teams can recover selected sessions, which become fully indexed in RUM for a 30-day analysis period with access to their events and attributes. The feature is intended to balance cost control with broader historical coverage, such as allowing a financial services team to investigate a customer-reported mobile trading display issue weeks after it occurred and confirm that the backend transaction was completed correctly.

Datadog is now available through Anthropic’s Claude Marketplace, allowing joint customers to purchase Datadog Agent Observability alongside Claude and apply part of their committed Anthropic spending toward it. Agent Observability helps teams monitor, trace, and evaluate Claude-based AI agents and applications, providing visibility into costs, reliability, failures, and output-quality regressions before they affect users. The marketplace offering builds on existing Datadog capabilities for tracing Claude API requests, managing Anthropic-related costs, and collecting Claude Code telemetry, while requiring no new technical integration or changes to current customer relationships with Datadog. Initially available to customers with US-based entities, the offering is intended to simplify discovery and invoicing, with broader availability planned over time.

Datadog’s Bits Investigation, launched in December 2025 as an always-on agent for investigating production alerts, has evolved from a complex multi-agent workflow into a simpler ReAct-based architecture that dynamically reasons over evidence and selects tools. By October 2026, newer models, improved routing, expanded data access, and code execution through the Datadog MCP Server had reduced average investigation costs from $25 to about $4 and average completion time from 11 minutes to roughly 2 minutes, while raising internal benchmark quality scores by 44% compared with the March 2026 release. The platform now processes and correlates data across sources more efficiently, enabling on-call engineers to spend less time gathering evidence and reach probable root causes sooner. Datadog is also developing Scout, a specialized SRE investigation model trained through in-house reinforcement learning, with plans to apply similar targeted training to autonomous remediation and faster service restoration.

Datadog’s C++ SDK extends Real User Monitoring to native C and C++ applications on Windows, macOS, and Linux, providing crash diagnostics, error tracking, and user-session context to help teams investigate failures beyond stack traces alone. It supports in-process crash handling for simpler deployments and Crashpad-based out-of-process handling for more comprehensive diagnostics, including all-thread traces and immediate reporting of issues such as stack overflows or heap corruption. Crash and handled-error reports can be correlated with RUM sessions, views, actions, network requests, operations, and long tasks, allowing developers to reconstruct the events preceding a failure and identify recurring issues by user impact. The SDK offers C++17 and C99 APIs, works with native build systems such as CMake, can be exposed through FFI bindings for other runtimes, sends data directly to Datadog without a local Agent, and includes consent controls for managing data storage and uploads.