Stephen Cihak: Harness: agents that deploy code

Why this conversation

Agents write code faster than people can review and ship it, so the bottleneck moves to delivery. Stephen Cihak is Field CTO at Harness. The host recorded the conversation on 2 November 2026 at SREday Amsterdam.

Harness, Stephen said, saw it coming a while ago: “there’s so much pressure with the agentic code that’s coming out that we really need to move to more of an autonomous SDLC.” That means agents at every stage: “you need to have agents plan the code, write the code, and agents that actually go out and deploy the code in a safe, secure, reliable way.”

Version three, at machine speed

They have built it over the past year and a half or so, “And just recently at Unscripted, we released our kind of version three of the Harness platform, which has autonomous steps in it”. The steps let agents deploy your code, and you choose how far each one may go. Stephen’s summary: “It’s a really good way to move away from deployment that operates at human speed to deployment that operates at machine speed.”

How much autonomy each step gets

The host asked how many levels of autonomy there are. Stephen walked up the scale. At the bottom, no agent at all: “I’m going to have a deterministic step that says, hey, no, you have to do this every single time.” Next, an agent that recommends, that you run the scan or do these things. Then a change advisor that puts a number on it, “there’s a 90 percent chance that this is OK to go”, with a human in the loop who says yes, or fails the build and sends it back for fixes. At the top, the agent decides: because the percentage is high and “blast radius is really small”, it moves forward on its own. “So you get to decide the level of autonomy each step in every stage”.

A data question, not a hunch

What does the change advisor look at? Policies you define. “So the size of the change, how complex of change is this in the code base? How many services does this touch? Is it part of critical infrastructure or not?” From those it can “basically give you a percentage-based decision matrix”, so you are informed about the change and the release by data. Above a threshold the agent goes ahead; below it, a human evaluates; or it only recommends, whatever the percentage.

The comparison Stephen drew is with the meeting most teams know, a decision taken on data “as opposed to a change advisory board”, where the developers come and describe the change and the service. With the change advisor, “it becomes more of a data question than a probabilistic kind of subjective answer that you were giving before.”

A new feature every 12 hours

The future is always hard to predict, Stephen said, but Harness is “drinking our own champagne”, and moving at a pace much faster than it has traditionally: “So we’re seeing a new feature being developed and released every 12 hours at this point.” Not a new model, Stephen corrected the host: “Well, it’s not a new model. It’s a new feature that’s within the Harness platform.” It can be as fully fledged as “a code coverage agent that can go out and automatically increase the code coverage for more critical areas of your code base”, or as small as a feature flag for an extra step. Harness, Stephen added, has just merged with Augment Code, and more and more capable models are coming, as Stephen sees it.

Getting started

“Actually, we have a free tier. You can go out today to harness.io and give it a shot.” It runs the latest version of Harness, and the first steps are done for you: “There’s AI agents out there that will create your first pipeline for you”, set up the repository and deploy it somewhere. For training and experimenting, the same advice: sign up for the free trial and try it.

When it comes out

The episode is edited and in the release order. Like every episode it goes out on YouTube first, then on Apple Podcasts, Spotify and Amazon Music.