AI Dev 26 x SF | Paul Everitt: The Shift to Agentic Engineering

Everitt / JetBrains / DeepLearning.AI — AI Dev 26 x SF: The Shift to Agentic Engineering (2026-05-22)

More code, fewer staff — the industry is on a bender. But what about quality?

At AI Dev 26 x San Francisco, Paul Everitt from JetBrains discussed the rise of agentic engineering and how old lessons can be adapted to build new professional practices.

(YouTube description, DeepLearningAI channel — AI Dev 26 x SF conference talk.)

A ~28:17 conference talk from the DeepLearningAI AI Dev 26 x San Francisco event, published 22 May 2026. Auto-generated English captions via yt-dlp fallback (the engagement-panel route timed out at 180s — recurring failure mode on ≥20-minute talks per §Acquire failure modes). VTT rolling-window dedup applied; 331 segments.

Speaker: Paul Everitt — Developer Advocate at JetBrains. Python old-timer (“first Python meetup in 1994”). JetBrains’ three-word pitch: “Privately held. Profitable. European.” — “25 years of doing this code intelligence stuff and developer productivity, we’ve seen some transitions before.”

The talk is the wiki’s first JetBrains-altitude practitioner call-to-arms on agentic engineering, joining the Ng keynote and the same-day Cline + Anthropic slide-gen talks in the AI Dev 26 SF conference cluster.

TL;DR

Five substantive contributions:

1. The problem framing — eight failure modes of more code, fewer people. Everitt opens with eight cited problems with the current AI-software-engineering moment, each pinned to a specific source:

#	Problem	Citation
1	Productivity gap. “It ain’t 10×, it’s 10%.”	DX study + Daron Acemoglu (Nobel 2024 economics) — “productivity is like five things in coding and in software engineering and coding is only one of those, and you could speed that up completely and it would still only be partial.”
2	Quality crisis — 50% phase-defect rates.	Live-from-the-conference “talk yesterday about defects, phase defect rates” + Simon Willison’s “challenger disaster” worry — “you take the human out of the loop, you let the agent shove stuff into production, maybe it drops a bunch of tables. No, I’m sorry, that was this week.” (Reference to the actual Replit-dropped-production-database incident.)
3	Price changes — “we can’t keep subsidizing these tokens. Guess what? It started to happen [this month].”	Ed Zitron newsletter + the register article cite.
4	Trust gap — “only 3% [in 2025] had a lot of confidence in the accuracy of the generated results.”	Survey cite (unnamed).
5	Token-maxing — employees gaming AI productivity metrics. “Uber is like, oops, annual gone in three months.”	Fortune magazine article cite — “a 67-point gap between how management feels and how engineers feel about this.”
6	Mental health / open-source maintainer PR-spam crisis.	Open-source community vantage.
7	Sovereignty / regulated-industry gap. “84% of Europeans are like, ‘Eh.‘” + insurance-industry / banking-industry skepticism on move fast and break things.	Live-from-the-conference Marina panel discussion + the Michael (insurance) dinner conversation.
8	Mega-layoffs at Atlassian / Block / Stripe / Amazon [this month, May 2026]. “Why do they do it? Their stock price goes up.” + Sam Altman’s “AI washing” counter-framing.	Direct citations.

Everitt’s frame: “You’ve been given the god box. What are you doing with the god box? Are you creating magnificent new solutions for your customers or are you just squeezing the profit margin?” The cited “27% of the work [at Anthropic] went to things that couldn’t have been possible without the god box” is the upper-bound counter-anchor.

2. The agentic engineering = build the thing that builds the thing reframe. Everitt’s central thesis is that “more code, fewer people” is the wrong frame; the right frame is “agentic engineering” — the OpenAI Harness Engineering article’s “we build the thing that builds the thing” + Simon Willison’s dark factory pattern (lights off in the factory because no humans are in it; humans are outside building the factory). The human is the bottleneck — “don’t eliminate the human, augment the human.” The reframe credits Karpathy explicitly: “Vibe coding created around 14-15 months ago by Karpathy and then he’s like ‘okay clean up — what I meant was agentic engineering.’ So he coined the phrase a few months ago and it seems to have caught on a little bit.” The reframe also explicitly invokes Grady Booch (creator of UML, “third golden age of software engineering” podcast): “isn’t this just software engineering? … coding was never really the big thing.” + the Dave Farley reference.

3. The nine-element agentic-engineering practice taxonomy — the central practitioner contribution.

#	Element	Worked treatment
1	Evals	”If the answer is button-go-clicky, it ain’t engineering.” + “we’re actually going to need data scientists doing actual work again because you don’t want the LLM to be the judge all the time.” Cross-references Sonar/Tom’s talk earlier in the conference.
2	Harness engineering	”From LangChain, a good Python company, and the thing — can you read it? — is that contrast: if you don’t own your harness you don’t own your memory. Bold provocative statement.”
3	Tooling / code mode	Anthropic + Cloudflare’s code mode — agent generates code, runs it in secure sandbox, “not sed and grep and walking around until it figures out the problem”. Pydantic’s Monty — Rust-based subset of Python for sandboxed tool-code execution.
4	Red-green TDD for agents	From Lenny’s podcast (Simon was on): “if you will write a broken test as the first step in your agent’s work, the agent will start to learn the way you like to write tests and will mimic your testing style. … it instead of wandering around trying to figure out how to please you, it knows exactly how to please you.”
5	Modularity for agents	”When you want to have parallel sub-agents and highly specialized sub-agents and really good context engineering, you might reorganize your code around some different principles.” Self-referential at JetBrains: “We’re JetBrains. We’re IntelliJ. We’ve been 25 years of this stuff. We got a big old code base.”
6	QA agents	From the OpenAI harness-engineering article: as humans become the bottleneck, “move more capabilities into the agent” — DevTools protocol / browser integration into the agent’s tool surface so it can collect its own instrumentation.
7	Observability	Pydantic LogFire pitched: “general system observability + AI observability — you probably need something that can do both.”
8	Orchestration	”500 Silicon Valley startups in series X of their funding” + Booch’s reminder: “remember your architecture.”
9	Context engineering + leadership & culture	Brandon at Unbox cited as the conference’s best context-engineering talk. FOBO (fear of being obsolete) enters Everitt’s vocabulary — “your folks are worried. You’re a leader. You have to get them going in the right direction, but you have to be honest and authentic.”

The taxonomy is the wiki’s first explicit nine-element practitioner-discipline-altitude treatment of agentic engineering as a discipline. Convergent with Osmani’s harness-engineering article (which focuses on the harness layer) but broader in scope.

4. The Grady-Booch Gang-of-Four call for agentic design patterns. The talk’s most-novel meta-contribution. Booch (creator of UML, co-author of Design Patterns: Elements of Reusable Object-Oriented Software) — “his cohort did unified modeling language, did design patterns, kind of built the field of modern software engineering. He knows there is a next thing called agentic patterns and he wants one of you in the audience to be the one that gets the ball rolling and helps us all figure out this discipline of agentic design patterns, agentic engineering.” Everitt closes by suggesting DeepLearningAI is the “home where we all come together and create this new discipline.” The wiki’s first named-Gang-of-Four-for-AI call — plausible single-source-deferred concept-page candidate: agentic design patterns.

5. The call to arms (not call to action) — engineers as the messengers who can reframe executive narratives. Everitt’s CTA “with a little bit of a twist”: engineers are uniquely positioned to reframe leadership’s more code fewer people narrative into engineering augmenting humans. “They’re good at what they do, but they know they can’t do your job, and so they listen to you. Let’s get them a better frame. Let’s get them back to engineering. … we should be doing agentic engineering and then they will start to see it mentioning everywhere and then they’ll start to believe and we’ll get them off of automating the current and replacing the current workforce and over to innovation augmenting the humans getting great bold new things done.” The argument is structurally parallel to Russell Wald’s academic-vs-industry transparency tension from the AI Index 2026 talk-track (5 days later): engineers and academics, respectively, as the credibility-bearing intermediaries who can recalibrate executive discourse.

Cross-references inside the AI Dev 26 SF conference

Everitt’s talk is unusually dense with cross-references to other talks at the same conference:

• “Andrew in the leadup to the conference in his newsletter, The Batch” — Ng’s framing, cited as the inspiration for the “we’re shaping the future right here, right now” opening.
• “In the talk just before from Sonar, Tom showed how all these models are different and they have different characteristics that change over time” — evals point.
• “Page stole this from me on the first day” + “Marina in her panel discussion” — sovereignty point.
• “Brandon gave one of the best presentations that I saw in the last couple of days” — context engineering point.
• “Mima [DeepLearning team helper]” + “the folks at DeepLearning” — closing dedication.

This makes the source a useful single-talk index into the AI Dev 26 SF conference programme — it situates the Ng keynote, the Cline / Anthropic same-day evals talks, and the AMD talk (which Everitt does not directly cite but which thanks Ng in identical language) inside a single practitioner narrative.

Why this matters in the corpus

Three sub-corpus roles for this source:

1. The wiki’s first JetBrains-vendor-altitude anchor on agentic engineering. JetBrains has not previously appeared as a substantive entity in the wiki. The 25-years-of-developer-tools + Python-since-1994 vantage is distinct from the model-vendor altitudes (Anthropic / OpenAI) and the IDE-extension-startup altitudes (Cline / Cursor).

2. The wiki’s first explicit nine-element practitioner-discipline-altitude treatment of agentic engineering. Convergent with Osmani (harness layer only) and Karpathy (the coining moment) but more taxonomically complete. Useful as anchor for the agentic-engineering concept page.

3. The wiki’s first Gang-of-Four-for-AI call attributed to Grady Booch. Plausible single-source-deferred concept-page candidate: agentic design patterns.

The W&W tagging (10 cells — comparable to Cline at 9 cells and Allen AWS London at 10 cells) reflects the talk’s reach across the eight-failure-mode problem framing (contextual/external-triggers + contextual/internal-barriers), the nine-element practice taxonomy (digital-seizing/* + digital-transforming/*), and the augmentation-over-replacement normative argument (strategic-renewal/business-model). The contextual/internal-barriers cell in particular picks up Everitt’s FOBO + 67-point-management-vs-engineer-gap observation, which is the wiki’s clearest practitioner-altitude characterisation of the trust-and-fear internal barrier to agentic-engineering adoption.

ASR notes

• Captions sourced via yt-dlp --write-auto-sub (engagement-panel route timed out at 180s — recurring for ≥20-minute talks).
• VTT rolling-window dedup applied — second-line-of-each-cue content kept, inline <c> per-word timing tags stripped, HTML entities unescaped (>> → >>), bucketed to ~5-second segments.
• Surface artifacts: “Duron Oamoglu” in ASR (correct: Daron Acemoglu); “Wilson” and “Willis” both appear for Simon Willison; “Grady BCH” (correct: Grady Booch); “Addi Osmani” (correct: Addy Osmani); “Karpati” (correct: Karpathy); “Sam Alman” (correct: Sam Altman); “Mackenzie” (correct: McKinsey).

Linked entities and concepts

Entities (already promoted, source_count bumped):

• DeepLearningAI — channel / publisher.
• Andrew Ng — direct citation.
• Daron Acemoglu — Nobel-2024-economics anchor on productivity-is-five-things claim.
• Andrej Karpathy — coined agentic engineering.
• Anthropic — code-mode worked example + 27%-counter-anchor.
• Simon Willison — direct citations (challenger disaster, dark factory pattern, what’s left for humans? so much stuff).
• Sam Altman — AI washing counter-framing.
• OpenAI — Harness Engineering article reference.
• LangChain — if you don’t own your harness you don’t own your memory citation.
• Pydantic — Monty + LogFire.

Dangling first-mentions (single-source, deferred per §Lifecycle author-entity promotion):

• Paul Everitt — JetBrains developer advocate.
• JetBrains — developer-tools company, European, profitable, privately held.
• Grady Booch — UML / Gang of Four co-author / third golden age of software engineering podcast.
• Addy Osmani — already promoted via 2026-05-15-osmani-agent-harness-engineering.
• Dave Farley — Modern Software Engineering author.
• Ed Zitron — newsletter author / AI-bubble critic.
• Brandon (Unbox) — context-engineering talk.
• Tom (Sonar) — prior talk on model characteristics.
• Marina — panel discussion.
• Kai — funny CEO-impression videos.
• Cloudflare — code-mode reference.

Concepts (last_confirmed bumped; substantive new content possible on agentic-engineering and ai-employment-effects):

• agentic-engineering — the central concept; nine-element practice taxonomy is the most-complete treatment in the wiki to date.
• agent-harness — if you don’t own your harness you don’t own your memory citation.
• ai-employment-effects — eight-failure-mode framing + mega-layoffs + token-maxing + FOBO + 67-point management-vs-engineer gap + Sam Altman AI washing.
• ai-washing — Sam Altman’s AI washing counter-framing on the more code, fewer people mega-layoff narrative (this talk is the wiki’s first carrier of the term; the BBC source supplies the dedicated labor-economist treatment).
• micro-productivity-trap — it ain’t 10×, it’s 10% + MIT 95% study + Uber-annual-gone-in-three-months.
• ai-benchmarks — we need data scientists doing actual work again + evals-as-first-element.
• automation-vs-augmentation — Everitt’s don’t eliminate the human, augment the human + dark factory framing.
• durable-skills — engineering-as-durable-discipline + Grady-Booch Gang-of-Four-for-AI call.

Source

• Raw transcript: transcript file (yt-dlp WebVTT → 5-second-bucketed segments; 331 segments; ~28:17 minutes).
• Public URL: youtube.com/watch?v=n366hY4JZ9U
• Channel: DeepLearningAI
• Conference: AI Dev 26 x San Francisco
• Published: 22 May 2026
• Caption track: auto-generated; no human-curated track at time of ingest.

Reading scope

Full ~28:17 transcript read end-to-end during ingest. Eight-failure-mode problem framing captured with citations. Nine-element practice taxonomy captured. Grady-Booch Gang-of-Four for AI call captured. Cross-references to other AI Dev 26 SF talks (Ng / Sonar / Marina / Brandon / Tom / DeepLearningAI helpers) captured as conference-programme index. Plausible single-source-deferred concept-page candidate: agentic design patterns.