The Pragmatic Summit 2026

The Pragmatic Summit — San Francisco, February 11, 2026

• Agenda

Gergely Orosz organized a one-day summit on AI and software engineering in San Francisco, hosted by Statsig. Attendees were selected through an application process — builders and practitioners. Sessions were recorded.

I’ve been a Pragmatic Engineer subscriber since ~ December 2021. This was the first Pragmatic Summit.

Sessions I attended

Welcome — Gergely Orosz

Gergely opened by noting he started Pragmatic Insights 8 years ago while at Uber, wanting to share engineering open secrets without hype.

How AI is reshaping the craft of building software — Vijaye Raji (CTO Applications, OpenAI), Tibo Sottiaux (OpenAI)

OpenAI’s engineering leadership described how bottlenecks keep shifting inside their own org: code generation got fast, so the bottleneck moved to code review, then to understanding customer needs.

• OpenAI reported a 5x productivity increase internally
• Tibo leads Codex with 33 direct reports in a flat structure
• Hiring ~100 summer interns/new grads — first batch. Codex handles onboarding
• Alexander, a PM, “hyper-leveraged himself with Codex” — building prototypes directly
• Designers ship code. Roles blur across product, design, and engineering
• When asked about a prediction for next couple years - replied “2 years is way too long” for predictions” answered with what he sees in ~ 6 months.
• Models are more capable when given larger tasks with built-in QA loops (feedback)
• Demo day depth has increased consistently — more corner cases solved, not just happy paths

Data vs. hype — Laura Tacho (CTO, DX)

Laura Tacho created the Core 4 metrics framework. 10+ years in developer tools and productivity. Her talk was around the space program — wonder balanced with pragmatism.

Industry numbers:

• 92.6% of developers use AI coding assistants; 44.1% use them daily
• 4.08 hours saved per week per developer on average (Google reports ~10% savings — hasn’t changed dramatically)
• 26.9% of code is AI-authored, up from 22% last quarter
• AI cuts onboarding time in half: ~90 days to reach 10 merged PRs (Q1 2024) down to ~40 days (Q4 2025). Effect persists 2+ years

Agentic adoption:

• 50.5% daily agentic tool usage
• OpenAI reported 1M Codex downloads in the prior week
• Trillions of tokens processed per week (per OpenAI)
• 95% of OpenAI developers use Codex (internal figure)

What the data actually shows:

• Some organizations saw 2x more customer-facing incidents after AI adoption. Others saw 50% fewer. Same tools, different outcomes. Tooling isn’t causal — engineering discipline and guardrails dominate outcomes
• “Orgs that were dysfunctional are dysfunctional faster”
• “Adoption doesn’t mean impact” — tools enhance individual productivity, not P&L performance
• Spray and pray doesn’t work. Successful orgs point AI at a concrete goal and measure progress
• AI measurement framework: utilization, impact, cost

Her recommendations:

• Invest in developer experience: feedback loops, documentation, fast CI, solid testing. “Just call it agent experience when chatting with the C Suite” 😉
• Treat AI as an organizational problem, not a technical one. Barriers are change management and lack of executive backing, not models or tools
• Reference: DORA AI capabilities model, ThoughtWorks AI readiness framework

Case studies cited:

• Haven (healthcare) — fine-tuned HIPAA-compliant model on 100K historical messages for migraine patient messaging. 3x industry average CSAT
• 18K Cisco developers using Codex
• JPMorgan’s multi-agent annotation framework (MAFA) for labeling customer interaction data

Reinventing software — Martin Fowler, Kent Beck, Gergely Orosz

Martin Fowler and Kent Beck in conversation. One of the strongest sessions of the day.

I hope to write more about this session in a follow-up post. For now, the key points:

• Fowler: “Programming is NOT the bottleneck”
• Fowler: “AI is an amplifier” — amplifies good practices and bad practices equally
• Fowler: “To go faster you need higher quality”
• Beck: “Value of software = features + what we can do in the future”
• Beck: “At this moment, no one has the answers”
• Fowler said he’s seen nothing “with the magnitude of AI” in his career — not object oriented, not the internet, not agile
• Beck: “What’s the smallest experiment I can run?” — the useful skill right now is validating cheaply
• Fowler compared this to a Venn diagram of agent experience and developer experience. The overlap was nearly complete: “Our craft done well, works for agents”
• Beck noted the temptation to go solo — one person managing agents instead of leading people. He pushed back on it - we are still building for humans.

Product engineering in an AI-native world

• Michelle Lim - Flint
• Tuomas Artman - Linear
• Drew Hoskins - Temporal
• Margaret-Ann Seger - OpenAI
  • Margaret’s blogs

Panel on how product engineering teams are changing.

• Artman (Linear): “It comes down to testing — as a code reviewer I concentrate on tests”
• “Product becomes engineer, or engineer becomes product” - lines blur
• PMs prototyping with Replit. Product teams using agents as standard workflow

What’s next: building world-class engineering orgs — Rajeev Rajan (CTO, Atlassian), Thomas Dohmke (ex-CEO GitHub), Gergely Orosz

Closing panel.

• Rajan: “Agents — we just shift to the right. New bottleneck.”
• Atlassian experimenting with smaller, more creative teams (managing down to sqrt(n) direct reports via agents)
• Dohmke: “Lines blur between product, designer, engineer”
• Dohmke Entire. First product is Checkpoints — an open-source CLI that collects agentic session information automatically, pushes to Git, enables rollback and rewind
• Gergely noted Entire’s funding as “the biggest seed round in dev tool history” (speaker claim) - Dohmke noted inflation 😏

Roundtable: AI product development — evals, annotations, tools

Small group session led by [RC Johnson] (Bumble)](https://www.linkedin.com/in/rcjohnson/).

• Two eval types: deterministic evals and LLM-as-judge
• Guardrails for off-topic detection using classifiers
• Scale challenges with live evals: latency and cost
• Product teams should own model selection and testing
• Reference: hamel.dev for evals guidance

Hallway conversations

• OpenAI Codex team: discussed agent repo strategies, skill composition patterns, and MCP vs. direct API tradeoffs
• Statsig: experiments and feature flags platform, dynamic config, positioning for AI-native teams

Sessions I missed (recordings to watch)

• Lessons from building Cursor — Sualeh Asif (CPO, Cursor), Alex Xu, Sahn Lam
• Vercel v0 and d0 agent — Malte Ubl (CTO, Vercel)
• New AI product at Ramp — Ian Tracey, Veeral Patel, Will Koh
• Coding agents for ICs — Simon Willison, Marcos Arribas (VP Eng, Statsig)
• Building AI applications — Chip Huyen
• High-performing teams — Nicole Forsgren, Gergely Orosz
• Uber’s agentic shift — Ty Smith, Anshu Chadha

Tools and products mentioned

• Codex (OpenAI) — agentic coding platform
• Cursor — AI code editor ($10B+ valuation)
• Statsig — think of it as a bit higher level than Launch darkly - product led experiments and feature flags
• Linear — issue tracking
• Entire / Checkpoints — open-source CLI for agentic session capture
• DX / Core 4 — developer productivity metrics
• DORA — AI capabilities model
• Replit — used by PMs for prototyping