AI Engineer Code Summit · Wed 11 Nov 2026 · Continental 4–6 (Mainstage)

Easier in public than in private

Ramp built its own coding agent, Inspect, and it now writes about three quarters of the company's code. Zach Bruggeman skips the plumbing for what came after: one shared agent for the whole company, every session public by default. Good practice spreads without training, and a session becomes the evidence that lets a reviewer trust a change.

Zach Bruggeman, Ramp · first 7 min of the talk · Links and further reading
The agent

Inspect, and why the technology isn't the story

Partial talk: the talk was cut short in the room, so the recording covers about its first seven minutes. His sidebar lists the sections: Intro, Kill silos, Build trust, Think bigger, Closing. This page covers the first two and the start of the third.

“As a quick introduction, Inspect is Ramp's collaborative coding agent.”00:01:26

About a year ago Ramp bet on building its own,

“with every session running remotely and being fully multiplayer.”00:01:26

It works alongside people in Slack, Linear and other tools, with design primitives, review agents and verification, and it's connected to the company's telemetry, databases and tools.

“Through this, we now see around seventy-five percent of all of Ramp's code. Being written and shipped by Inspire.”00:01:26

A year ago the technology was the news: sandboxes in under five seconds, internal context wired up, agents working multiplayer. Not any more.

“the technology isn't nearly the barrier that it once was.”00:01:26

So the talk is about

“the lessons we've learned around building a culture around inspect”00:02:55
From the notes: “inspect is Ramp's coding agent, multiplayer”. His demo shows it: a session started from a Slack message about greyed-out approve buttons, with the chat, a live browser and the agent's trace side by side. The technical deep dives are on the links page.

Evidence: transcript 00:01:26, 00:02:55

Lesson one

Kill information silos: one Inspect for everyone

“One thing about working at Ramp is that it really makes you start thinking about everything through the lens of saving time and money.”00:03:29

If everyone in a company of hundreds or thousands spends fifteen minutes setting up their own agent,

“Once you start multiplying that across an organization, this is days lost for what is very rote work.”00:03:29
“You realize this setup cost isn't just one time.”00:03:29

Best practice keeps moving, and the knowledge spreads unevenly. An individual might count that as an edge; a company can't.

“When you're trying to win, you need everyone on your A game.”00:03:29
“Build tools that kill information silos.”00:04:25

When a pattern works, they push it to everyone at once. People can still tune their own,

“But Inspect is always available as a very solid baseline.”00:04:25

and they've never had to mandate it. By design there's only one:

“So part of the design of Inspect is that there's just one single inspect.”00:05:14
“You just keep using Inspect as you normally would, and benefit from the improvements that everyone's making over time.”00:05:14
From the notes: “build tools to kill information silos”. The reading: the silo isn't data, it's know-how: how each engineer has set up their agent. A shared agent makes that a company asset that improves for everyone at once, the opposite of every engineer maintaining a private harness. For a client: how many agent setups do they have, and who updates them when best practice changes?

Evidence: transcript 00:03:29, 00:04:25, 00:05:14

The spine

Easier to work in public than in private

They don't run weekly training sessions on using AI.

“With every Inspect session being public within the company by default.”00:05:38
“It's actually easier to work in public than in private.”00:05:38
“And you can learn how to use it by watching how your teammates use it.”00:05:38

The agent helps:

“You can literally ask inspects to go through their recent sessions and tell you about how they work, how they structure prompts.”00:05:38

His slide shows the answer for one colleague: he starts with the outcome and says how he'll verify it; he points at prior art, a similar PR or session; and he splits big problems into one session per risky piece. The Inspect team learns from it too:

“We frequently see the way people use inspect in ways that we didn't originally imagine.”00:05:38
From the notes: “you can inspect other people's sessions”. The reading: public sessions turn everyone's work into training material, and the agent can summarize it. The three habits are worth copying as they are, and they match Lovable's outer loop and Diogo's “stages that can be verified”.

Evidence: transcript 00:05:38

Lesson two

Building trust: the session is the artifact

“It's all about building trust.”00:06:44

Better models helped, but much of this year has been learning to trust what an agent builds.

“So much of building trust in tools comes from learning from other people.”00:06:44

Usually that means conference speakers, and that's a narrow sample.

“Instead of grouping around a select few leaders, can we build systems that inherently share the knowledge of everyone?”00:07:27
“It's a living artifact of how that change came to be, when a pull request goes up.”00:07:27

A reviewer sees the conversation, the tools and specs used, before-and-after screenshots, even a video of the verification agent clicking through the app. The recording ends here.

From the notes: “i had to cut that talk short”: the rest of “Build trust”, all of “Think bigger” and the closing aren't captured. The reading: it's the same move as your Cognition note this morning, “they are sharing the videos etc to show that things work”: evidence travels with the change. And it answers Diogo's root of slop, unchecked output that pushes the checking onto someone else; here the author's agent does the checking and attaches the proof. For a client: what does a reviewer see beside the diff?

Evidence: transcript 00:06:44, 00:07:27