Intel / article

Bits investigations are faster, more accurate, and cost less

Source: datadoghq.com
Date: Published Nov 2, 2026

Context

Datadog says its Bits Investigation , which investigates production alerts to find their root cause, now costs about $4 and takes about 2 minutes per investigation on average, down from $25 and 11 minutes at its December 2025 launch. The company's post attributes much of the change to simpler architecture. The first version ran a coordinated set of agents through dedicated phases: gather , query telemetry, then generate and test root-cause hypotheses. A later version switched to a simpler design known as ReAct, in which an agent reasons about the evidence, picks a tool and uses the result to decide what to check next, which brought the average cost to about $6.50.

The latest step lets Bits write and run code through the Datadog MCP Server, so a single tool call can query several data sources, correlate the findings and return a summary. Processing the data before it reaches the model means the model handles less context and needs fewer round trips. Datadog also credits newer models and better routing between them. Its quality claim uses a different baseline: a 44% higher score than its March 2026 release on an internal benchmark that compares Bits's findings with the known root causes of past incidents. All of these figures are Datadog's own measurements. The post ends by pointing to Scout, Datadog's model trained for site-reliability investigations, which it announced separately and says matches leading models on investigation tasks at lower inference cost.

Key takeaways · AI-distilled

Datadog says Bits Investigation's average cost per investigation fell 84%, from $25 at its December 2025 launch to about $4, and average run time fell 82%, from 11 minutes to about 2, as of October 2026.
The first Bits ran a pipeline with dedicated phases for context retrieval, hypothesis generation and testing. A simpler ReAct loop, where each agent picks its next tool from the evidence, plus newer models and tools, cut average cost to about $6.50.
The latest version adds better and code execution through the Datadog MCP Server: Bits writes code that queries several data sources and summarizes the results in one tool call, so less raw data reaches the model and fewer round trips are needed.
Datadog scores quality by comparing Bits's findings with the known root causes of past incidents, and reports a 44% higher score than its March 2026 release. The benchmark is internal, so the gain is vendor-reported.
Datadog is training Scout, an SRE investigation model built with its in-house reinforcement learning system on a wide range of failures. It says Scout matches leading models on investigation tasks at lower inference cost, and plans to apply the approach to remediation.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
model routing — Sending each request to a model chosen by the difficulty of the task, rather than using one model for everything.

Read the source www.datadoghq.com