AI-Wiki

The Future of AI Agents with Andrew Ng (LangChain Interrupt '26)

9 min read

The Future of AI Agents with Andrew Ng (LangChain Interrupt ‘26)

Andrew Ng sat down with Harrison Chase for a wide-ranging fireside chat at Interrupt, the agent conference by LangChain. Watch the full session and get insights on: how fast coding agents have evolved and how the way small teams ship is being reshaped; the “product management bottleneck” and why accelerating software development just moves the constraint somewhere else; how enterprises are getting incremental AI wins but missing the bigger transformation; the data architecture rework most companies haven’t started yet … and so much more.

— LangChain YouTube channel (Interrupt ‘26 fireside, hosted by Harrison Chase)

A ~32-minute fireside chat at Interrupt ‘26, LangChain’s agent conference, between Andrew Ng (founder of DeepLearning.AI, AI Fund, AI Aspire; co-founder of Coursera) and host Harrison Chase (CEO of LangChain). Ng’s second year on the Interrupt stage. The conversation spans coding agents, the future of software-engineering teams, enterprise AI adoption, vendor optionality, open-weight models, and the unstructured-data rearchitecture — most of it from the operator chair of someone running multiple AI companies and advising Fortune-50/G2000 businesses through AI Aspire.

TL;DR

1. Coding agents moved fastest; the rest of the hype/doom over-shot. Ng: the hype and the doom (the “job apocalypse,” which “I don’t think is going to be a thing”) both got more traction than he expected; on the positive side, coding agents took off faster than he’d have guessed. The “everything changes every three months” cliché is mostly false — except for coding agents, where it feels true. His own mix shifted from “almost all Claude Code” six months ago to a rotation of Claude Code + OpenAI Codex + Gemini CLI + open-source coding agents, “changing rapidly,” including coding from his phone against a Mac Mini in his office.

2. The product-management bottleneck became an everything bottleneck. A year ago Ng wrote about the product-management bottleneck: when building gets fast, deciding what to build (scoping, customer feedback, prioritisation) becomes the constraint. Over the last year it got “much worse, in a good way” — and when software gets 10–100× faster, “pretty much everything else becomes a bottleneck”: a marketing bottleneck (marketers can’t keep up describing all the new features), a legal-compliance bottleneck (a week for legal sign-off is fine on a 3-month build, painful on a 1-day build), a design bottleneck, and so on.

3. Small teams of high-context, empowered generalists. Ng increasingly sets up teams of 1–10 people, “high-context, highly empowered generalists” given wide guardrails to run inside — writing and shipping code and drifting into marketing copy, terms-of-service first drafts, design, etc. The pigeonhole argument: if a team of two humans must cover five functions, each human must play more than one role — and AI makes them “slightly less bad” at the roles they’re weak in. Most people who do this well come from a deep engineering background, but he’s seen product managers, marketers, and operations people learn to build effectively too.

4. Building blocks / LEGO bricks + keeping agents current (Context Hub). Ng’s mental model for the future of software engineering: a proliferation of building blocks (AI ones — RAG, agent frameworks, evals, guardrails — and non-AI ones — UI components, auth, databases). Developers who master enough blocks combine them combinatorially (“exponentially as a function of the number of LEGO bricks”). The friction: coding agents don’t know the newest blocks (a model’s knowledge cutoff predates e.g. nano-banana’s API). His fix is Context Hub (Rohit Prasad’s — “a Stack Overflow for AI agents” serving the latest docs/SDKs to agents; note: name-collides with, and is distinct from, LangChain’s own Context Hub), which loads current docs so the agent makes the API calls he’d rather not memorise.

5. Education is changing — what to learn and how. What developers must learn has shifted (coding agents, building blocks, a bit of product management). On how, Ng thinks the transformation of education has been over-hyped and “isn’t quite here yet,” but is iterating: DeepLearning.AI launched CodeDream.ai / LearnDream.ai, where instead of watching a course you have a conversation — a simulated one-on-one video call with an AI Ng you can interrupt, with interactive JavaScript-instead-of-video demos you can type your own prompts into. Newer courses (e.g. a just-launched Transformers course) lean on interactive visualisations over passive video.

6. Enterprise adoption: bottom-up isn’t paying off; the transformation needs top-down. Via AI Aspire (his advisory firm with Chris Tann, advising Fortune-50/500/G2000), a consistent theme: everyone invested in bottom-up “thousand flowers” innovation, and “for the most part it is not paying off” — boards are asking where’s the ROI? Keep doing bottom-up (it generates ideas and real incremental efficiency), but it yields point solutions and incremental gains, not transformation. The loan-underwriting example: automating the one-hour human loan-approval step is a small efficiency gain; rethinking the whole workflow into a “get approved in 10 minutes” product is the transformation — and that takes someone with broad scope to redesign marketing, routing, diligence, and execution end-to-end, i.e. a top-down motion complementing the bottom-up one.

7. Cost savings vs. growth; swing-for-the-fences vs. incremental. Cost savings are fine but capped; growth has almost no ceiling, so Ng pushes for growth ideas (the 10-minute loan, automated/augmented contact centres and drive-through ordering → more delightful service → growth). On ROI: “I wish I knew” — measuring ROI is as hard as measuring business. A counterintuitive lesson: driving incremental gains can be harder than transformative gains (you can’t get people to “work 50% harder,” so a 20–50% target forces creative solutions). Businesses send AI Aspire spreadsheets of 300+ ideas; narrowing to a small portfolio of thoughtful bets with meaningful resources takes hard technical and business analysis, and tends to need a top-down resource-allocation motion (the cost of prototyping has plummeted, but you still can’t fund everything).

8. Forward-deployed engineers: a good thing, slightly over-hyped. FDEs are having a Silicon Valley moment; Ng likes them but expects most firms to have many in-house engineers and a smaller embedded FDE team. Building agent workflows is genuinely hard — it needs business understanding, customer-facing skill, observability/evals, change management — “deep technical judgment.”

9. Optionality and vendor lock-in. Because “the leading AI model rapidly changes” and he has “no idea what would be the leading coding agent a year from now,” optionality is very valuable. Ng’s personal practice: almost never sign longer than a one-year contract regardless of the 20–30% discounts vendors offer for three-year deals. He values vendor-neutral tools (names LangSmith) for observability and preserving optionality, and supports open-weight models (his teams use them, sometimes fine-tuned; they sit ~6–9 months behind the frontier but are cheap enough to matter) — voicing concern about White-House noises on inspecting models before release as “a war on open source/open weight.”

10. Data strategy before agents — the coming unstructured-data rearchitecture. A very common enterprise pain point: rethinking the data architecture. The last 10–20 years organised structured data (tables, relational); now AI can process unstructured data (text, images, PDFs, audio, video), and getting it to agents “at the right time, in the right place” is suddenly much more valuable. Ng has found no good off-the-shelf solution and is running internal experiments to rearchitect his own unstructured data. He predicts “tens to hundreds of millions of dollars” of data-rearchitecture projects across many businesses over the next few years — the blockers being fragmentation, governance, no consensus schema, data on laptops, and permissions designed for humans not agents (does an agent inherit my permissions?). A coding aside: he prefers NoSQL (MongoDB) for rapid prototyping because redesigning a relational schema mid-iteration is “so annoying,” moving to relational/scalable stores only at large production scale.

How this source touches the wiki (dynamic capabilities)

  • digital-sensing/digital-scouting — Ng’s continuous scan of the coding-agent frontier (the shifting Claude Code / Codex / Gemini CLI mix), the vendor landscape, and what’s working vs. not across the G2000 via AI Aspire.
  • digital-seizing/strategic-agility — the optionality discipline (≤1-year contracts, vendor-neutral observability, open-weight hedging) and the top-down resource reallocation behind a thoughtful portfolio of bets.
  • digital-seizing/balancing-digital-portfolios — narrowing 300+ candidate ideas to a small portfolio of high-conviction bets; swing-for-the-fences vs incremental allocation; cheap prototyping but finite funding.
  • digital-transforming/redesigning-internal-structures — small teams of high-context generalists with wide guardrails; the 10-minute-loan end-to-end workflow redesign as the unit of transformation.
  • digital-transforming/improving-digital-maturity — DeepLearning.AI’s reskilling toward building-blocks mastery and product sense; the CodeDream/LearnDream interactive-learning push to raise workforce digital maturity.
  • strategic-renewal/business-model — drive growth, not just cost savings (10-minute-loan product, contact-centre-as-growth) and the unstructured-data rearchitecture as the enabler of agent-era value creation.
  • contextual/external-triggers — the coding-agent adoption wave that “took off faster than I would have guessed” as the external trigger reshaping software work.
  • contextual/internal-enablers — getting the data strategy right before building agents; embedded forward-deployed engineers; executive scope to drive top-down redesign.

Roles override (roles: explicit): ceo, cso, cdo, cto, transformation-lead. The source’s centre of gravity is enterprise AI strategy + how to organise (engineering) teams + where the transformation ROI actually is — the C-suite-strategy and transformation-ownership roles, plus the CTO for the software-engineering-org claims. The override drops the broad finance/HR/marketing cluster the eight cells would otherwise inherit.

Linked entities and concepts

  • Concepts this source informs: enterprise-ai-adoption (bottom-up vs top-down; the loan example; ROI; data rearchitecture), agentic-engineering (coding-agent rotation; building blocks), automation-vs-augmentation (the product-management/everything bottleneck; generalist multi-role teams), durable-skills (domain judgement + scoping as the scarce input), warner-wager-process-model / dynamic-capabilities (the enterprise-transformation reading).
  • Entities (already in wiki): Andrew Ng, Harrison Chase, LangChain, DeepLearningAI.
  • Dangling (single-source mentions, deferred per the second-source promotion rule): Chris Tann (AI Aspire co-founder; first appearance), Rohit Prasad (Context Hub; first appearance), AI Aspire / AI Fund (Ng’s firms; concept-mentions), Context Hub / CodeDream.ai / LearnDream.ai / LangSmith (products named once — promote on a second source).

Source-to-source relationships

Source-quality note

Transcript fetched via the youtube-transcript-skill (Playwright/DOM route); the video carries human-curated (manual) English captions (kind: manual) alongside an ASR track, so transcript fidelity is high. The fetch returned the transcript doubled (the body repeats from ~0:00 a second time); the duplicate tail was disregarded at ingest. Transcript provenance does not feed confidence per [§Lifecycle].

Relationships

Supports

  • Agentic coding and persistent returns to expertise — Ng's *product-management bottleneck* and *high-context generalists* are the practitioner-side picture of the AEI's measured division of labour — humans decide *what* (planning, scoping, judgement), agents do *how* (execution). Both conclude the scarce, value-bearing input is domain judgement and the ability to decide what to build, not implementation throughput.
  • How DBS Bank Makes Everyone an Innovator (Leaders at All Levels, Ep. 9) — Ng's central enterprise claim — *bottom-up 'thousand flowers' innovation generates point solutions and incremental efficiency, but the transformation needs a top-down motion to redesign the whole workflow* (the 10-minute-loan example) — is the diagnosis that DBS's playbook model operationalises: top-down transformation KPI (20% of scorecard) meeting bottom-up delivery (journeys, agents). Two enterprise-transformation sources naming the same bottom-up/top-down complementarity.
  • Architecting AI-Native Organizations: How to Redesign Work at Scale — Joe Beutler — Both argue AI value comes from *redesigning work at scale*, not deploying tools task-by-task. Ng's small teams of *high-context, empowered generalists given wide guardrails* who cross functional boundaries is a concrete org pattern for Beutler's *architecting AI-native organizations* thesis. Shared cell: redesigning-internal-structures.
  • London Executive Forum 2026 — A leader's guide to advanced team structures in an agentic world — Both address how teams restructure around agents. Allen (AWS) traces the builders-to-orchestrators role-shift and end-to-end W&W operationalisation from the vendor chair; Ng supplies the practitioner version — generalist multi-role teams of 1–10, the product-management-bottleneck-becomes-everything-bottleneck dynamic. Shared cells: redesigning-internal-structures, strategic-agility.
  • Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness — Optionality and vendor-neutrality from two angles. Ng's practitioner discipline — *never sign longer than a one-year contract regardless of discount, because I don't know which model/coding-agent will lead in a year* + open-weight support + a vendor-neutral observability layer (LangSmith) — is the buyer-side counterpart to Kilpatrick's supply-side claim that model capability erodes harness/provider lock-in over time. Both treat the leading model as a fast-moving target that punishes lock-in.

Graph View

Backlinks