Observability as an AI control plane

Datadog is trying to turn the software that watches systems into the software that supervises artificial intelligence agents. On March 9, 2026, it said its new Model Context Protocol server would give artificial intelligence agents secure, real-time access to unified observability data inside Datadog. (datadoghq.com) Observability used to mean collecting traces, logs, and metrics after something broke. Datadog’s pitch now is that the same telemetry can sit in the loop while agents make decisions, so the system watching the factory floor also becomes the foreman. (datadoghq.com) That shift only works if the platform can see each step an agent takes. Datadog’s large language model observability product traces inputs, outputs, latency, token use, errors, and intermediate steps across agent workflows, which is the raw material you need to judge whether an agent is fast, cheap, and behaving. (datadoghq.com) The company has been building that layer for more than a year. In June 2025, Datadog said new large language model observability features would monitor agentic systems, run structured experiments, and evaluate both custom and third-party agents inside one product. (datadoghq.com) Then it started adding workers on top of that data. On June 10, 2025, Datadog introduced three Bits AI agents for operations, development, and security, which turned Bits AI from a chat assistant into software that investigates incidents and proposes fixes. (datadoghq.com) The clearest example is security. Datadog says Bits AI Security Analyst can autonomously triage Cloud Security Information and Event Management signals, deliver investigation results to Slack or Jira, and recommend verdicts in minutes instead of forcing analysts to dig through alerts by hand. (datadoghq.com) On March 23, 2026, Datadog put a number on that claim. The company said Bits AI Security Analyst can cut investigations that used to take hours down to as little as 30 seconds and reduce threat investigation time by up to 98 percent. (datadoghq.com; seekingalpha.com) Datadog is making the same argument on the product side, not just the security side. After buying experimentation company Eppo in May 2025, it launched Datadog Experiments on April 2, 2026 to connect product tests with business metrics, product analytics events, and application observability in one workflow. (datadoghq.com; datadoghq.com) That matters because feature releases and agent changes now look similar: both are controlled rollouts that can go wrong in production. Datadog’s Feature Flags product, launched February 3, 2026, ties each flag to observability data so teams can automate rollouts and rollbacks, while Experiments measures whether the change improved the business outcome you actually wanted. (datadoghq.com; datadoghq.com) The executive sell is no longer “buy more dashboards.” It is “use one telemetry system to watch agents, test changes, enforce guardrails, and prove results,” which is also the line analysts have started to pick up as Datadog’s artificial-intelligence-driven growth story. (investing.com; datadoghq.com) Datadog’s own filings show why it wants that story to land. In its 2025 annual report, the company said revenue reached $3.43 billion and described itself as a unified observability and security platform, which gives it a reason to push observability from an operations budget line into a broader control layer for artificial intelligence, security, and product teams. (stocktitan.net; sec.gov) The next fight is whether customers accept one vendor as the referee for all of this. Datadog’s partnerships with companies like Sakana AI and its integrations across OpenAI, Amazon Bedrock, and other model stacks show the company knows the control plane only works if it can plug into the systems where agents already live. (datadoghq.com; datadoghq.com; datadoghq.com)