Datadog has Claude Code and Codex build a Redis-compatible server and a Kafka-compatible streaming engine, and leans on test harnesses instead of code review
Datadog's engineering blog describes 'harness-first engineering': instead of reading every line of agent code, the team invests in automated checks that say quickly whether the code is correct. A single Claude Code agent (Opus 4.5) built redis-rust, a Redis-compatible server, and the team added layers of checks as bugs slipped through: a shadow-state oracle, deterministic simulation testing with fault injection, TLA+ specs, Kani proofs, Maelstrom/Knossos linearizability checks and the official Redis Tcl suite, then shadow traffic against Redis 8.4 in staging. For Helix, a Kafka-like streaming service on object storage, multiple agents (mainly Claude Code and Codex) worked from written invariants and design records, with simulation runs scaled to 10 million seeds.
The numbers (as reported)
| redis-rust memory | 87% reduction after agent-guided optimisation (started at 8x Redis 8.4's memory) |
|---|---|
| redis-rust latency | Comparable to Redis in production-like staging |
| Helix simulation | 500 seeds per component, then 10 million seeds across all components |
| Helix throughput | About 93% of peak disk throughput |
| Spec overhead | Writing invariants takes roughly 2-3x the effort of writing the code |
Facts
- Company / subject
- Datadog
- Tools
- Claude Code, OpenAI Codex
- Models
- Claude Opus 4.5
- In-house tools
- pup
- Codebase
- greenfield
- Work type
- New product
- Stack
- Rust, TLA+, Kani, Stateright
- Published
- 2026-03-09
- Last checked
- 2026-11-02
What the source claims Practitioner (own blog/paper)
Claimed by the builder, writing about their own work. Not checked independently.
- redis-rust reached production-like staging with comparable latency and an 87% memory reduction after agent-guided iteration -87% memory
“redis-rust reached production-like staging with comparable latency and an 87% memory reduction after agent-guided iteration”
- Producing to Helix averaged 22.2 ms against 116 ms for the baseline Kafka cluster 22.2 ms vs 116 ms
“Producing to Helix averaged 22.2 ms, compared to 116 ms for the baseline Kafka cluster.”
Caveats
Company writing about its own internal projects; redis-rust reached staging, not full production. First post in a series, so later results may differ. Datadog sells observability, which the approach leans on.
More stories by tool: Claude CodeOpenAI Codex
Related topics: New products built with agentsGreenfield projects