How Datadog Built a Universal Machine Tool for Claude Code

Channel Anthropic

Date May 6, 2026

Duration 31 min

Tags Datadog, Claude Code, Enterprise, Tool Framework, Engineering Platform

TL;DR

Datadog reached 90% AI coding tool adoption across their engineering org in four months, with Claude Code handling two-thirds of that usage. But as sessions grew more ambitious, the tools being generated per-session became unmaintainable. Sesh Nalla (VP Engineering) explains how Datadog built Temper — a constrained framework that makes tools secure, reusable, and composable across the entire org.

Key Takeaways

90% adoption in 4 months — Datadog's engineering organization adopted AI coding tools faster than any other internal tool in company history
Tool sprawl is a real problem — As Claude Code sessions grew more ambitious, engineers were generating one-off tools per session that didn't compose, weren't reusable, and had inconsistent security postures
Temper: a constrained tool framework — Datadog's answer is a framework that enforces security constraints, promotes reuse, and makes tools composable across sessions and teams
Security is the core constraint — Temper enforces least-privilege access: tools can only do what they declare, and declarations are reviewed before a tool enters the shared library
Compound value over time — Unlike one-off session tools, Temper tools accumulate; the more tools in the library, the more capable Claude Code sessions become
Verification is built in — Temper tools ship with test cases; Claude can verify tool behavior before relying on it in a session

Summary

The Adoption Curve

Sesh Nalla opens with a remarkable stat: 90% of Datadog's engineers adopted AI coding tools for production work in the four months after launch, with Claude Code driving two-thirds of that usage. This wasn't mandated — it was demand-driven. Engineers saw peers shipping faster and wanted the same. The pace of adoption created a new problem: the tools being generated to support those sessions.

The Tool Sprawl Problem

As Claude Code sessions became more ambitious — spanning multiple services, requiring Datadog-specific API access, integrating with internal monitoring — engineers started generating custom tools per session. A tool to query the metrics API. A tool to parse log formats. A tool to trigger deployments. Each session generated its own version. None composed. None were tested. Several had security issues — overly broad API access, credentials in tool code, no audit logging.

Temper: The Solution

Temper is a constrained framework for building Claude Code tools that are secure, reusable, and composable. The constraints:

Capability declarations — Every tool declares what it can access (APIs, services, data) before review
Least-privilege enforcement — Tools can only access what they declare; the framework rejects over-scoped tools at review time
Test coverage requirement — Every tool in the shared library ships with test cases that verify correct behavior and error handling
Versioning and changelog — Tools are versioned; breaking changes require a new version, not an in-place update

The Compound Effect

The breakthrough insight behind Temper is that tools should compound. A one-off session tool is sunk cost; a Temper tool is an investment. Every new tool added to the library makes every future Claude Code session at Datadog more capable. After six months, Datadog engineers starting a new session can invoke 200+ battle-tested, security-reviewed tools covering their entire internal stack.

Results

Since deploying Temper, Datadog has seen session length increase (more ambitious tasks), security incident rate from Claude Code decrease to zero (vs. several per month with ad-hoc tools), and tool reuse rate of 78% (tools are being used, not regenerated). Engineering teams are contributing tools to the shared library proactively.

Notable Quotes

"We had 90% adoption in four months. That's incredible. And it created a tool sprawl problem we didn't anticipate — because 90% of engineers generating tools means a lot of tools, and most of them were one-offs."

"The key insight behind Temper is that tools should be an investment, not a cost. If you generate a tool and use it once, you've wasted it. If you generate a tool and it gets used 10,000 times across the org, you've compounded."

"Security isn't a post-hoc layer. It's a constraint at the design level. Temper makes it impossible to ship a tool that violates least-privilege — the framework rejects it."