Notable verbatim quotes

Speaker labels are absent — these are all Tammuz Dubnov unless otherwise noted. Quotes are verbatim from transcript.md including speech-to-text artifacts; intended meanings are in square brackets only where essential.

On the definition of "AI-native"

[§2 — definition]

"For me, being AI native. Means that the person that cares, the person has the authority to make the decision. It's also the person who can do the work. That basically AI collapses the gap. Collapses the handover."

[§2 — what AI-native is not]

"the answer or the second answer I do get is, well, being an animated [AI-native] means that rpms are designers in our QAs, more people who've been pull requests. And that's somewhat true. But in my side, that is the symptom."

On the real bottleneck

[§3 — handover, not code-gen]

"the bottleneck for a while now has not been how fast can I write cod. E? The bottleneck is doing that they hand off."

[§3 — compression to minutes]

"It means graphs and stuff can be days or weeks to 13 minutes."

On wrong ways — Uber & Microsoft

[§4 — Uber 6× spend]

"Uber has increased the area [AI] spend since 2024. By 6x. Now that 6x growth was completely used within four months, they are out of AI budget for the rest of the year."

[§4 — CFO conversation]

"costs are getting part of justified. People are more and more tokens. But me as a business leader, I'm not seeing how there's a link to actually moving faster. On features."

[§4 — Microsoft Claude Code rollback]

"Microsoft. Kind of similar phenomena, they get really cloud code [Claude Code]. And now they're taking it away. Because just giving hot code to more and more people in your organization is a wonderful way to earn money. It is not the right way to speed up the order."

On where the tokens go

[§5 — $18 of $100]

"from. $100 that you spend on the only $18 actually goes to meaningful code that actually gets shipped to users."

On Shopify as positive example

[§6 — 50% acceptance]

"they said that 50% just gets accepted. It means that the quality is so high. That the work getting done by those non-engineers can actually emerge as is"

On wrong-way patterns

[§7 — disinterested devs]

"developers to work on tasks that they're not interested in the agentic era. They'll just hand it off to an agent. And they will let the agent check itself ... Huge token consumption in a single session."

[§7 — prototyping tools are "hurry up and wait"]

"you will get pms and designers prototyping tools so they can prototype faster ... But that doesn't solve the big bottleneck of cup but handover. It just gets them to create a signal faster. And wait for the next for longer."

[§7 — Claude Code for everyone = PR fatigue]

"getting into PMs and designers is a great way. For them to go to your desks every other day to say, hey, help me set up the environment. ... build the PRs that are thousands of lands and total nonsense. And basically ways for them to increase the PR fatigue that your death [dev] team is already feed. Ing."

On harness engineering

[§8 — definition of harness]

"harness. Means that you saw the agent make mistake and you make it unfeasible for the agent to make the same sticky game [mistake]. You put in some sort of wall."

[§8 — harness must evolve]

"the harness needs to adapt. The mistake is made. Mistake is realized. The harness needs to adapt to prevent that mistake from happening again. So it's not something that's supposed to be handheld. It's something you're supposed to grow and evolve organically."

[§8 — overconfident agent]

"There's nothing worse than an overconfident agent."

On the failed PR — authority boundaries

[§9 — the storage-architecture PR]

"the developer said, oh, we can't store this until we have an internal discussion, decide how we want to store them. It's going to be in the db or bang [bucket] in the bucket. ... there's no way we're going to go with what the agent decided"

[§9 — expectations]

"the expectations shouldn't be, hey, 100%. If a pm opens a PRO be mad as a developers for not merchants, you want them to be that part."

On measurement — the headline numbers

[§10 — 74% merge rate]

"an average non-technical contributor in our organization ... they're merging is about 74% Which means that one out of four, they accidentally overstepped. And that's okay. That's not PR fatigue."

[§10 — 84% zero-dev-touch]

"we measure from PRs that are open and merge. How many get merged without any dev interfering without them pushing more commits to fix change adjust. And again for us that is 84%."

On the moral

[§11 — democratising authorship]

"The point of being handed [AI-native] is actually democratizing authorship so that the person that is accountable, person that has the authority is a person who can do the work. Without independent on a bunch of other individuals in the org and hoping that they do the work or waiting for them to do the work."

From the Q&A

[§13 — feature flags + dev autonomy]

"they let the developer make those decisions. We depend pretty heavily on feature types [feature flags]. So you say developer, go for it, make all your product decisions. Everything is feature flag. Why? Because we need to be able to merge it quickly."

[§14 — not opinionated]

"the harness is super lean to adapt your practices."

[§15 — risk + size labels on every PR]

"we now have an agent in every token. It labels the risk level and the size. So you as an operating nine [opening a PR] ... You will see that three of those are low risk small. You will see that one is like extra large."