Influential practitioners

AK Andrej Karpathy

• Sequoia AI Ascent 2026 talk — christened the shift from vibe coding to Agentic Engineering. Named December 2025 his inflection point. "You can outsource your thinking, but you can't outsource your understanding." The New Stack
• Coined "context engineering" as successor to prompt engineering. Tweet, June 2025
• The viral "Karpathy CLAUDE.md" — Karpathy observed the failure modes; Forrest Chang distilled them into a copy-pasteable file (Jan 2026) that crossed 110k+ stars. Source · our position (use the file; ignore the unverified accuracy claim).

SW Simon Willison

• "Agentic Engineering Patterns" — open guide, first chapters Feb 23 2026.
• Pragmatic Summit talk (Feb 2026) — "Engineering practices that make coding agents work."
• Defines an agentic tool as one that can both generate and execute code.

AR Armin Ronacher

• "Agentic Coding Recommendations" (June 2025)
• "Tools: Code Is All You Need" (July 2025)
• "Things That Didn't Work" (July 2025)
• "A Year Of Vibes" (Dec 2025) — switched defaults to TypeScript and Go because agents do better there.
• "A Language For Agents" (Feb 2026)

MH Mitchell Hashimoto

• "Agentic Engineering in Action" (Zed blog) — "I endeavor to always have an agent doing something at all times."
• Popularized "harness engineering": when an agent makes a mistake, build a test/validation/lint that lets it self-check next time.
• Recurring voice in Zed's "Agentic Engineering" series — see the interwoven workflow for our position on the framing.

BC Boris Cherny

• Latent Space: Claude Code deep walkthrough
• Pragmatic Engineer: Building Claude Code
• Lenny's: "What happens after coding is solved"

Design philosophy: latent demand, prototype-driven instead of PRDs, plan mode is a design document — never let Claude write code until plan is approved.

GL Geoffrey Litt

• "Code like a surgeon" (Oct 2025) — stay in the loop on primary work; delegate secondary tasks.
• Long-running theme: malleable software — end users adapting tools via AI.

SY Steve Yegge (Sourcegraph)

• "Cheating is All You Need" (2023, foundational)
• "Revenge of the Junior Developer" — agentic coding will surpass chat coding.

HH Hamel Husain

• "LLM Evals: Everything You Need to Know"
• "Your AI Product Needs Evals"
• AI Evals for Engineers & PMs on Maven (3000+ students from OpenAI, Anthropic, Google).

Core claim: 60–80% of dev time on real LLM products should be error analysis, and most failed LLM products share no robust evals as the root cause.

EY Eugene Yan

• "Patterns for Building LLM-based Systems & Products" — the canonical seven-pattern reference: Evals · RAG · Fine-tuning · Caching · Guardrails · Defensive UX · Collect feedback. Organized on a data ↔ user × defensive ↔ offensive 2×2. See Patterns · Yan's 7-pattern map for our cell-by-cell position; the framework also drove three new pattern entries in our playbook (Guardrails, Defensive UX, Feedback flywheel) and the LLM-as-evaluator bias notes in Evals.
• Co-author of "What We've Learned From A Year of Building with LLMs".

sw swyx · Shawn Wang

• Latent Space — long-form interviews with most major figures in agentic coding.
• "Scaling without Slop" (2026) — coding agents "break containment and eat the world."

LM Lance Martin (LangChain)

• "Context Engineering for Agents" (June 2025) — taxonomy: Write / Select / Compress / Isolate.
• Cites Manus AI data: a typical agentic task uses ~50 tool calls — context budget is the real constraint, not model capability.

DB Drew Breunig

• "How Long Contexts Fail" — names the four context rot modes.
• "How to Fix Your Context" — Pruning, Summarization, Offloading.

WY Walden Yan (Cognition / Devin)

• "Don't Build Multi-Agents" (June 2025)
• "Multi-Agents: What's Actually Working"

IS Ilya Sutskever

• Dwarkesh Podcast — "We're moving from the age of scaling to the age of research" (Nov 25 2025) — frames 2012–2020 as research, 2020–2025 as scaling, 2026+ as research again. Another 100× scaling would help but not transform.
• "Jaggedness." Models pass the hardest PhD-level exams but lack the taste, judgment, and reliability of a human expert. RL on benchmark distributions creates models that are superhuman at test-taking but unreliable in practice. → Patterns · Jaggedness
• Continual / on-the-job learning as the next paradigm. Rejects one-time pre-training; AGI as a system that learns after deployment.
• Timeline: 5–20 years for a system that learns as efficiently as a human.