Your agents write more code. What happened to your DORA metrics?

November 12, 2026 @ 1:00PM EDT

Coding agents write more of the code, and most teams feel faster. Few can show it. In GitKraken’s survey of 554 developers and engineering leaders, 84% said AI makes them more productive, and only 20% could measure it. DORA’s 2025 research adds the harder half: AI now lifts throughput, but delivery instability keeps rising.

DORA metrics can tell you that delivery changed. They can’t tell you which agent, tool or repo changed it, or what your agents were missing when work stalled. Jeremy Castile of GitKraken and May Walter of Hud take agent-led delivery apart one question at a time.

GitKraken Insights shows what DORA can’t: which developers get real value from agents, which tools and models earn their cost, and which repos are ready for agents at all. Hud, the runtime code sensor, adds the context no repo holds. It checks every PR against how the functions it touches run in production. It compares each function to its own baseline as the rollout takes traffic. And it gives the agent writing the fix the evidence from inside the function that failed.

Agenda:

What happened to your DORA metrics: what DORA and GitKraken research say about speed and stability.
What DORA can’t tell you: adoption and autonomy, tool and model cost, and repo readiness.
Where agents run out of context: review queues, regressions that reach users, and recovery that starts from logs.
Three rollout practices and the metrics each one affects.

Takeaways:

A scorecard that pairs DORA metrics with adoption, autonomy and cost per output.
How to tell which tools, models and repos are holding delivery back.
How to set a baseline before changing your process.
How seeing runtime behavior next to the diff can make review faster.
A staged path to automatic rollback: recommendation first, policy in your CD later.