I Found Out How Datadog Verifies Everything Claude Code Writes - Meet Temper!

Hi, I'm Shii-chan! Today I found an exciting story about how Datadog verifies the huge systems that Claude Code helps them build. Let me walk you through it!

Claude Blog claude.com

What was announced?

The Claude Blog published an interview with Sesh Nalla, an engineering lead at Datadog. The piece is titled "How Datadog built a 'universal machine tool' for Claude Code." It's about how Datadog built Temper, a system for verifying the code that AI agents like Claude Code generate inside the company.

Sesh puts it this way:

"You're no longer writing the code; you're shaping the work."

Instead of writing code directly, engineers are now shaping the work itself, he says.

Why it matters

Sesh points out that while AI agents have gotten great at generating code fast, the effort needed to verify and operate that code has actually grown. The speed of verification hasn't kept up with the speed of generation.

So Datadog split the job in two: agents write a specification for the control logic, and a separate system checks whether that specification is actually correct. That system is the kernel called Temper.

What changes

Here's how Temper works:

• An agent writes a specification
• The kernel runs four independent layers of verification on it
• Once verified, the specification runs as-is

The division of labor is clear: agents write specs, and the kernel compiles them. That means correctness gets guaranteed at the specification level, without engineers having to review every line of generated code by hand. Control logic is written as explicit state transition tables, which is the key idea here.

Dive Deep

Temper's verification runs across four layers:

• Symbolic reasoning proves that guard conditions and invariants hold
• Exhaustive state exploration visits every reachable state
• Deterministic simulation surfaces edge cases through fault injection
• Randomized property testing runs around 1,000 pseudo-random action sequences

Each feature is also expressed through three contracts:

• Behavior contracts: state transitions and safety properties
• Data contracts: entity types and operations, in machine-readable form
• Authorization contracts: default-deny policies

Datadog has a track record here: a distributed queuing system called Courier in 2024, an evolutionary optimization harness called BitsEvolve in September 2025, and a Kafka-comparable streaming service called Helix. Temper is the culmination of that work, serving as the agent control plane, the tool-building layer, and the lifecycle API. On Helix, Sesh notes:

"In a few days we had a fully functional Kafka comparable system"

So a Kafka-comparable system came together in just a few days.

Sesh compares all this to machine tools in manufacturing:

"Machine tools are the jigs, fixtures, gauges, and mills...They were the breakthrough of industrialization as parts became composable, inspectable, and replaceable."

He sees the same shift happening in software now, as parts of the system become composable, inspectable, and replaceable.

Wrap-up

• Datadog built a kernel called Temper to verify code generated by AI agents like Claude Code
• Agents write specifications, and the kernel verifies them across four layers: symbolic reasoning, exhaustive state exploration, deterministic simulation, and randomized property testing
• Each feature is described through Behavior, Data, and Authorization contracts
• Temper builds on earlier projects, Courier, BitsEvolve, and Helix, and now sits at the core of how Datadog runs its agents
• This one is especially worth reading for engineers and platform teams trying to figure out how to systematize verification when working with AI coding agents