The Pragmatic Summit

I left the inaugural Pragmatic Summit this week with the overall impression that while the capabilities of AI tooling are still advancing organizational bottlenecks are quickly becoming the real constraint. Generating code isn’t the hard part anymore. Startups without much process procedure and bureaucracy are thriving in this ambiguity and large orgs are having a much harder time.

I’ve never before been to a conference where it felt like every attendee I talked to could be on stage. The hallway conversations were all excellent and some of them have stuck with me longer than what I heard from the presenters.

I’m annotating my notes and personal impressions here and may be taking things out of context. If you want the full unfiltered detail subscribe to the Pragmatic Engineer (you should just subscribe anyway) and watch the content. Every single one of these sessions is worth the time investment.

How AI is Reshaping the Craft of Building Software (Tibo Sottiaux Head of Codex, OpenAI Vijaye Raji CTO of Applications, OpenAI)

It felt like their main goal was not exposing too much about how OpenAI operates or saying anything that might grab headlines. This was in stark contrast to hearing Simon and Sualeh later in the afternoon. I guess OpenAI is all grown up now.

• Spreading knowledge quickly in Slack and in showcases was my main takeaway from this talk (I’ve already scheduled AI dev tools office hours at work)

• Similarly to Lauras talk can hear how they are measuring with help of ai tools and using measurements to drive roadmaps and feed backlogs

• The most talked about item from this whole section was not how and what they are building (a lot of the attendees are already at Yegge’s stage 8) but the fact that Tibo has 30 direct reports

The State of AI in Engineering (Laura Tacho, CTO of DX)

This is the type of talk I’ve become used to seeing as a regular attendee at SRECon. I’ve officially become that voice of “Its just an opinion until we measure it”. I’m glad to see that the speed of AI code generation is forcing organizations of all sizes to wake up to this reality.

• The hot off the presses numbers: 92% of devs using AI assistants, ~27% of code AI-authored (up from 22% QoQ), average 4hrs/week saved (~10% productivity gain)

• The uncomfortable split: some orgs see 50% fewer incidents, others see 2x more. “If you’re dysfunctional already, AI makes you dysfunctional faster.”

• Adoption ≠ impact. High adoption, low transformation. Steep dropoffs from pilot to production.

• “Time to tenth PR” I love this as an onboarding metric. Can be applied to internal team/project moves too, not just new hires

• Her framing: stop chasing lunar colonies, build gas stations. Aim for Mars but solve problems on Earth.

• DevEx matters more than ever. DORA AI capabilities model as a readiness framework. Set goals, measure progress — spray and pray never worked and it definitely doesn’t work here.

Lessons from Building Cursor (Sualeh Asif Co-founder, Cursor)

Nobody is as refreshingly candid as a first-time founder of a wildly popular company. He only seemed to hesitate occasionally when he realized he was pre-empting a product launch or revealing actually valuable internal insight about how they build things.

• Released their own model (between Sonnet and Opus 4.5), used RL to teach it to actually use grep and search effectively

• Infrastructure at their scale can’t be bought — orchestrating millions of sandboxes is its own unsolved problem

• Cloud agents are only 1% of their compute today. Can’t scale 1000x without fundamental UX improvements. “Caring about security isn’t enough to drive this product.”

• Monitoring agents is nothing like monitoring services — natural variance from minutes to days, no clear signal for degraded vs. working hard. Mentions Temporal and Restate as workflow engines worth watching.

• Self-summarization at compaction checkpoints for longer tasks — dump context for the next session to grep. He thinks the tool needs to do this its not just a workflow problem.

• The “self-driving codebase” vision: allocate $100K/day, split usage across security, tech debt, bugs, backlog. Human review eventually becomes optional for certain classes of work.

• “Coding got solved in 6 months” — expect 1-2 more capability jumps in the next year. ICs need to develop manager instincts to adapt.

Proving Code Works (Simon Willison)

Putting Simon, Chip Huyen and Martin Fowler head to head just tells you all you need to know about the quality of presenters at this conference. I fully plan to go back and watch both of those other talks when I find some time. I’m glad I picked Simon, his energy and enthusiasm is infectious as always, and him reminiscing about Django’s founding in Lawrence (rock chalk) reminded me that I have been reading his blog for more than 20 years.

• His 4 stages of AI adoption: chatbot Q&A → agents writing code → agents writing more code than you → you don’t read the code anymore

• The trust inflection: Opus 4.5 was the first model that earned his trust for consistent results on classes of problems

• Red-green TDD is back and it’s fun now. Tests are free to maintain. “No longer even remotely optional.”

• Conformance-driven development and his new tool Showboat. Agents build markdown proof they exercised the API

• “Our job is to deliver proof, not code”

• Use GitHub PR reviews to quiz the agent “where did you construct the SQL query I should look at”

• The lethal trifecta for prompt injection: private data + malicious instructions + exfiltration vector. Cut one leg.

• Cookiecutter templates for agent consistency. The same principle as human teams applies first implementation has to be perfect, next 5 people copy it, agents do this faster

• Skills erosion is a choice. The unreliable teaching assistant is actually better because it forces skepticism.

• Mentions new tool Rodney (Go-based browser automation via rod) as a better alternative to Playwright/Chrome DevTools MCP

Leading Eng Teams in the Age of AI (Nicole Forsgren, Frictionless author)

Nobody can interview techies like Gergely can (yes, looking at you Lex Fridman). Every time he was on stage the presenters were always steered to reveal some compelling insight that they wouldn’t have revealed otherwise. I left this talk really excited to read Nicoles’ book and fighting the urge to march around my office with a drum chanting “Don’t forget the DORA metrics!”

• Productivity measurement was already bad, now it’s worse. “Vibes isn’t a metric” and leaders who trust their gut for humans don’t trust agents on gut feel either.

• Human bottlenecks (code review, release candidacy, security reviews) were already constraints; AI amplified them

• Cognitive load research: 3-4 hours of deep work is human max. Faster feedback loops from AI increase cognitive load rebuilding mental models dozens of times in 30 minutes.

• SPACE framework: satisfaction, performance, activities, collaboration, efficiency/flow

• Change control with agents is an emerging unsolved problem

• Adoption metrics expose other org problems, but you have to dig in

• Give explicit permission for 10% experimentation time. Psychological safety: “Will I be fired for using this tool if it makes mistakes?”

Recurring Themes

• The gap between individual productivity gains and organizational transformation is the defining challenge right now.

• Devex is the leverage point for humans and agents alike. If you want to sell devex improvements just call it agentx.

• Proving code works (tests, conformance, documentation) matters more than writing it. It probably always did…

• Monitoring and observability for agents is an unsolved problem

• The “self-driving codebase” concept is coming and nobody is sure what it means for engineering orgs in 2026 but everyone building tools sees it coming

Nicole said it best: “Agents need to see and understand the system so humans need to be able to see and understand the system first.” We’ve all got virtual assistants that are just as happy to get to work fixing documentation and solving long-standing devex issues as they are working on code. Let’s use them to make life easier for the humans while they are still in the loop.