How I came to this work.

About three years ago, after more than twenty years in CEO seats running software companies, I finished my last role and finally had time on my hands. I found myself completely absorbed by the generative-AI revolution that was just starting – reading and analyzing almost every piece of writing that came out, and not only on the technical side. The social, historical, political, and economic implications fascinated me at least as much as the models themselves. The whole shape of what AI was about to do to the world had me in a state I hadn’t been in for years.

But the further I went, the more one specific itch kept returning. I had always believed that software development was the area where AI’s future would be most visible first, because it was already the most advanced area of AI usage. If I really wanted to understand where this was going, I needed to build. There was just one problem: I had not written production code in a very long time. Different languages, different runtimes, a different operating model entirely. I felt completely out of shape. But the thought kept returning – why can’t I do it?

So one day I just sat down and dove in. The first move was small: experience what it felt like to work inside an agentic IDE. Within hours I was addicted. That experience became the first platform I built, AI-assisted from end to end – letting me move on things I hadn’t touched in years.

The second platform, just one year later, was already purely agentic: I directed intent and reviewed output; the code wrote itself. By that point I had forced myself through every serious tool in the space, and the agent harnesses layered on top of them, comparing them on real work rather than on benchmarks. I was running multiple agent sessions in parallel and routinely generating hundreds of thousands of lines of production code per month, operating in a mode that had no analog in how software was built even two years earlier – let alone two decades. In the gap between those two builds I watched the shift from 2024 to 2025 happen in real time. The whole way software gets built was changing under me. That was when I felt AI-pilled.

I use the word addicted deliberately. The pull is the same one that hooked me on programming in the first place, decades ago – you type something, the machine obeys, and a small jolt of it worked, a hit of dopamine, lands before you’ve thought about what comes next. Agentic coding turns that hit into a slot machine: you write a sentence, the agent disappears for a minute, and comes back with a working feature, or a near-miss that’s one more prompt from working. The reward is variable and the friction is gone, the same loop that makes a video game hard to put down. Every win suggests three more things to try, and the cost of trying is a sentence. I lost more than one night to just one more prompt, and I was nowhere near alone.

When Adam Fisher from Bessemer Venture Partners approached me about this project, I saw an opportunity to study the same shift from the other side of the table – across their portfolio and beyond, at scale – and to put two perspectives in the same frame: the hands-on experience I had been accumulating as an AI-native builder, and twenty-plus years of running software companies. The combination let me ask sharper questions than either viewpoint alone would have allowed. This playbook is the result.

Across our conversations with portfolio companies, four patterns kept appearing. Some teams believed they were making progress but had no benchmark against which to measure it. Others were moving quickly but kept their methods close, making it difficult to compare approaches. Some were spending heavily on AI without seeing the investment translate into output. And some knew what needed to change but could not get the organization to move. We realized the field needed more than another report: it needed a practical guide for founders, CEOs, engineering leaders, and boards navigating the transition to agentic engineering.

The work behind this document

During the first half of 2026, we spoke in depth with CTOs and engineering leaders from more than twenty companies, including Ramp, Lemonade, Wonderful, and DriveNets. They ranged from small AI-native startups and mid-stage software companies to late-stage organizations with codebases over a decade old. The study spans consumer software, B2B SaaS, infrastructure, security, e-commerce, marketing, and several other verticals – broadly enough that the patterns are not specific to any one industry.

The interviews ran in successive rounds, with findings synthesized between each. The framework barely needed to change as the sample grew round over round – itself a finding. The underlying patterns are more consistent across companies than the surface differences suggest.

This playbook’s production followed the same agentic model. The thesis, interpretation, editorial judgment, and final prose were human; agents helped analyze interview evidence, cross-check external research, and challenge drafts through several models. The source lived in HTML and Git rather than Word, allowing agents to work in parallel, propose changes as diffs, preserve a complete version history, test layouts, and generate web, PDF, and slide artifacts through code. This made iteration across research, writing, design, formats, and versions far faster than a traditional document workflow, while human judgment determined what the evidence meant and what merged.

Why a playbook, not a report

This is a playbook, not a report. A report tells you what other people are doing; a playbook tells you what to do next. The engineering leaders we spoke to did not want a trends deck – they wanted to know: where am I on this curve, what am I missing, what is the next move? The four-level maturity scales in each step are the answer. Read them honestly, locate yourself, identify the gap, start building. The field examples scattered through the asides are not the prescription; they are calibration points.

We deliberately included real stories from the field. Every leader we spoke to was curious how their organization was doing relative to the others – the most common question we heard, in some form, was “am I behind, ahead, or about typical?” The anonymized vignettes in the asides exist to answer that question without naming companies. They also do something a framework alone cannot: they generate ideas. More than once, a leader asked about a play another company had run, and went off to try a version of it the next quarter. Stories travel where frameworks don’t.

What surprised us

What surprised us most through the interviews was the pace. We had expected to find a few aggressive movers and many cautious laggards. Instead we found something stranger: every leader we spoke to had made moves that twelve months ago would have been considered reckless – week-long org-wide tool cutovers, 60% management cuts, PM-to-engineer ratios inverting, $1M AI bills exceeded in two months. And every one of them was making those moves with conviction. The pace alone tells you something.

We came away convinced of one thing above all: the technical changes are real and moving fast, but already well ahead of most organizations’ ability to absorb them. Many mistake that motion for absorption. Track lines of code as proof you’re transformed and you’ve taken a placebo: the number climbs while nothing real changes. The decisive work over the next two to three years won’t be just about the tools and infrastructure. It will be about the people who have to change how they work, moving from writing code to directing agents, and about the organization that has to change shape around them: the team structures, planning cadences, and ownership models that decide whether the new speed compounds or gets absorbed. The three problems interlock: once the infrastructure is real and enough people convert, organizational redesign stops being optional. This is the agentic awakening: a wave moving through the software industry, overturning methods and practices that worked for decades. Some already see its shape; most are still trying to understand what it means for their work, their careers, and their companies. That is the arc of this playbook: build the infrastructure, convert the people, restructure the organization.

Liran Eshel · June 2026 @liranesh