Chase, Sproul & di Vittorio — The Future of AI Agents: What Will Interrupt 2027 Look Like? (LangChain Interrupt 26)
Harrison Chase, Brace Sproul, and Caroline di Vittorio kicked off Day 2 of Interrupt 26 with an inside look at where AI agents are headed, the unveiling of LangChain Labs, as well as the latest features in LangSmith Fleet.
— channel description, LangChain
TL;DR
A ~22:10 LangChain Interrupt 26 Day 2 keynote (manual English captions, 193 segments), published 2026-05-21 on the LangChain YouTube channel. Three speakers: Harrison Chase (co-founder and CEO; segments 1–10 + closing), Brace Sproul (LangChain; segments 11–13 product walkthrough), and Caroline di Vittorio (LangChain; segment 13 live Fleet demo). The framing device: imagine Interrupt 2027 — what would the topics be a year from now?
The substantive contributions are six.
1. The two-types-of-agents typology (~1:22–2:42). Chase predicts a divergence — already emerging — in the agents being built:
| Type | Characteristics | Examples |
|---|---|---|
| Long-horizon agents | Run for minutes / hours / days; code execution, planning, sub-agents, multi-agent systems, skills; outcomes and goals as ways to extend the horizon | Knowledge-work agents that do more and more valuable work over time |
| Customer-experience agents | Latency-critical; voice (and maybe video) as modality; brand-sensitive; talking to an end user | Customer support, sales, end-user-facing |
“There is a shared stack underneath. There are also differences in terms of the technology. And I think one of the big questions that we’ll be thinking about over the next year is how common is this shared stack versus how particular are the technology pieces that you need for each?” This is the wiki’s first structural typology of agent-product surfaces distinct from the frameworks/runtimes/harnesses/no-code Build-layer typology of Chase’s ADLC piece — the new axis is what the agent runs against (the world / time horizon vs end-user-experience), not where the agent lives in the build stack.
2. Voice agents — pipeline vs native (~2:42–3:58). Today’s voice pipeline = speech-to-text → agent (operating in text space) → text-to-speech sandwich. Emerging = native speech-to-speech models (“OpenAI released V2 of theirs two weeks ago” — i.e. around 7 May 2026). “While I would say the consensus today seems to be for applications where you really care about having control, they’re not quite steerable enough yet, we do expect that to change.” The open question for Interrupt 2027: pipeline / native / combination.
3. Every agent needs a sandbox (~3:58–4:47). “We think all agents will need a sandbox, especially these long-horizon style agents. So coding is really good for a variety of tasks. It’s not just writing software — it’s for data analysis, it’s for web browsing, it’s for image gen, for deep research.” The framing anecdote: “when they think about building agents for their marketing team, one of the ways they think about it is they think about giving that marketing team a software engineer. What would that software engineer build? What apps would it build to make the marketing team’s job easier? That’s what giving an agent the ability to write and execute code is.” LangChain launched sandboxes “yesterday” (Interrupt 26 Day 1).
4. The rise of open-source models — three drivers (~4:47–6:01). (a) Capability: “the performance of these base open models without any type of post training for particular tasks is already approaching that of frontier models” — benchmark gap closing fast; (b) Cost: “in particular for coding agents, these agents are burning through a ton of tokens really, really fast — open models offer a cheap alternative”; (c) Trainability for particular domains: “they can be trained for your particular domain… as companies build up a lot of these traces, a lot of these agent runs, how can you use those to improve the model over time?” — directly setting up the continual-learning section.
5. Agent identity & auth patterns (~6:01–7:26). “How do agents take actions? On whose behalf do they take those actions?” Two emerging patterns Chase observes — both will coexist:
- Act-on-behalf-of-user: agent uses the calling user’s credentials; “if I’m accessing an agent and that agent has access to Slack and I tell it to look up something in Slack, it will use my credentials and it’ll have access to whatever I see in Slack. If Julia or a colleague uses it, they might get a different answer because it might see different things.”
- Fixed-service-account: “whoever is interacting with that agent, they will always use that fixed set of credentials, so they’ll see the same responses.” Started with OpenAI’s GPTs pattern; “we started to see actually some SaaS providers make it really easy for agents to create their own accounts so that they could have their own fixed set of credentials.”
The forward-looking claim: “being really, really precise about when to use which one and making that clear to users will be important.” This is the wiki’s first articulation of agent-auth-pattern as a structural product-design choice, not just an engineering detail.
6. Continual learning across three layers (~7:26–10:42) — the conceptual centerpiece. Chase articulates a three-layer model of any agentic system, each independently improvable:
| Layer | What it is | Examples |
|---|---|---|
| Model | The base foundation model | Sonnet, GLM4, GPT-4 |
| Harness | ”The code surrounding the model that connects it to the environment” | Deep Agents, Claude Code, pi |
| Context | ”What we provide to the harness as ways to guide it on particular tasks” | agent.md, skills |
Continual learning worked examples Chase names from stage:
- Model layer: Ramp + Prime Intellect fine-tuned Qwen 3.5 to be really good at Ramp sheets; “latency is very, very low and the accuracy is very, very high.” The example of why open models matter — they’re fine-tunable.
- Harness layer: MetaHarness (MIT + Stanford) — “used an agent to optimize a coding harness… outperforms human-written harnesses… it wasn’t changing the model at all, it was just editing the harness.”
- LangChain’s own: “we moved from top 30 on terminal bench two to top five just by changing the harness itself. So no changes to the model, only changes to the harness, and we saw a big increase in performance.” The third primary-source confirmation of this exact claim, following LangChain Engineering blog (10 March) and the broader pattern from OpenAI Codex (11 Feb).
The classical-ML analogy (~9:20–10:01) — the conceptual move that ties it together:
“In classical machine learning, you have the model, you have the training data, you do some gradient descent, and that updates the weights of the model. When you’re updating the agent in general, depending on the layer you’re at, if you’re working at the harness or context layer, it’s not exactly gradient descent, but the evals that you write act as a forcing function. So in the example of the MetaHarness… those evals are providing a similar type of this training gradient. And so evals and traces are incredibly important for this learning.”
This is the wiki’s clearest single articulation of evals-as-gradient for non-model layers — a load-bearing framing that ties the ADLC’s eval phase to the continual-improvement substrate of every other phase.
LangChain Labs and Fleet (~10:42 onward). Chase announces LangChain Labs as a research group inside LangChain “aimed in particular at continual learning”; “LangSmith, with all these traces already in there and all the feedback associated with it, this is a really, really solid foundation for doing this type of continual learning.” Brace Sproul then walks through LangSmith Fleet updates: 200+ built-in tools; first-class Arcade partnership for 7,500+ additional tools; MCP support; native Slack / Gmail / Outlook channel integration; sharing model like Google Docs; auth management; cost tracking and usage controls; first-class human-in-the-loop; built on deep agents open-source harness; downloadable agent files for code modification.
Internal-LangChain operational metrics (Caroline di Vittorio’s live demo, ~13:40–19:05). LangChain’s own go-to-market agent (built in Fleet without code, integrated with Salesforce / BigQuery / Slack / Gmail) reports:
- 84% of the go-to-market team uses it weekly
- Lead-to-qualified conversion up 240%
- 40 hours saved per rep per month on average
The agent was “originally built by an engineer directly in code, but when we built Fleet, we rebuilt this agent directly in Fleet so the go-to-market team could own this agent’s implementation entirely end-to-end without having to write a single line of code.” The wiki’s first concrete “engineer-built-it-in-code → rebuilt-it-in-no-code so the domain team could own it” migration with quantified outcome metrics.
Closing operational note: “We actually added a free model in [Fleet]. For a limited time there’ll be a free model — and backing up what we were talking about earlier, it’s an open-source model powered by Fireworks AI, one of our fantastic partners here.” The product release is congruent with the open-source-models-rising thesis from the keynote’s first half.
Caveats. Vendor-CEO keynote on the company’s own conference channel; every metric is vendor-cited. The 84% / 240% / 40h GTM-agent claims are unaudited internal LangChain operational data — verifiable only if LangChain publishes the methodology. The two-types and three-layer typologies are predictive frameworks Chase explicitly flags as “I’m going to be wrong in a lot of these, this is going to look really silly and really stupid” — treat as forward-looking commitments, not empirical observations. The Top 30 → Top 5 claim was already on the wiki via Trivedy in March; what’s new is Chase repeating it as a vendor-CEO stage claim.
Why this matters in the corpus
This source lands three structural extensions to the wiki’s existing LangChain corpus:
- The Chase / LangChain anchor doubles. Prior wiki: one Chase source (ADLC, 9 May) + three other LangChain-affiliated sources (deep-agents-harness Feb 17, Trivedy March 10, Y Combinator-channel cross-references). This Interrupt keynote is Chase’s second substantive source — triggers entity-page promotion per the second-source rule. LangChain entity: 4 → 5 sources.
- The three-layer continual-learning model (model / harness / context) is the wiki’s clearest single articulation of what continual learning of an agentic system actually means, complementing Zhang’s academic formalisation at the vendor-CEO altitude. Together they bracket the continual-improvement-of-the-agentic-system thesis on the academic and product-vendor altitudes.
- The Top 30 → Top 5 LangChain claim now has a third primary-source confirmation — Chase’s stage assertion at Interrupt 26 on top of LangChain blog (10 March) and the cohort of corroborating harness > model claims (Karten & Zhang, Lee et al. MetaHarness, Pan et al. NLAH). This makes it the strongest single-claim convergence in the wiki’s harness > model corpus — three independent primary-source citations of the same numerical magnitude.
Pair with CS153 (one day earlier). Both Chase and Tan articulate layered-system decompositions of agent thinking, landing within 24 hours of each other in late May 2026:
| Source | Axis | Decomposition |
|---|---|---|
| Tan / CS153 (20 May) | Agentic-primitives → company-structure | Skills = employees / Resolvers = org chart / GBrain = internal process / Check-resolvable = audit / Trigger evals = performance reviews |
| Chase / Interrupt 26 (21 May) | Learning-systems → agentic-system-layers | Model / Harness / Context — each independently improvable, with evals-as-gradient |
Both founders extend the everyone-builds-agents corollary: Hu’s AI founder / IC / DRI org structure (24 April) + Chase’s domain experts will be building agents, not just giving feedback (this source). The 2026 founder-CEO doctrine is converging on decompose-the-agent-stack and let-domain-experts-own-each-layer as the operating posture.
What was actually ingested
The full ~22:10 transcript was read end-to-end. Manual English captions are available alongside auto-generated; the youtube-transcript-skill picks the manual track when present, which gives noticeably higher transcript quality than the YC Root Access yt-dlp fallback batches earlier in this session. Speaker transitions are clearly marked in the transcript (Chase → Sproul at ~17:38 via “Now, I’m going to hand it off to Caroline” — though actually Brace hands off back to Caroline at 17:46; minor mid-keynote handoff confusion noted in the body).
Two ASR-style spellings to flag (despite manual captions): “LinkedIn” at 0:41 is almost certainly “LangChain” (Chase saying “one of the things that we think a lot about at LangChain — it’s actually in our mission and vision statement”) — likely a transcription mis-edit or punctuation artifact. “GLM4” (referring to the Zhipu/THUDM open model) and “Qwen 3.5” (Alibaba open model) are spelled correctly. The chapter start_ms: 0 fields in raw frontmatter remain a known YouTube-metadata bug — the build script’s chapter-interleaving parses the start: string instead.
Linked entities and concepts
Entities promoted by this source:
- LangChain — already entity (4 sources); bumps to 5 with new section on the Interrupt 2027 keynote and the three-layer continual-learning model.
- Harrison Chase promoted from Dangling to entity page — this is his second substantive source (first: ADLC, 9 May). New entity page filed.
Dangling — single-source mention, deferred:
- Brace Sproul — LangChain (product walkthrough segments 11–13). First wiki mention.
- Caroline di Vittorio — LangChain (live Fleet demo segment). First wiki mention.
- LangChain Labs — new research group announced; first wiki mention.
- LangSmith Fleet — already named on the LangChain entity page; this source quantifies its operational metrics (84% weekly usage on internal GTM agent, etc.).
- LangSmith Engine — referenced in the description’s “extra resources” link; not substantively discussed in the talk body.
- Fireworks AI — open-source model partner powering Fleet’s free-tier model. First wiki mention.
- Arcade — tool-integration partnership (7,500+ tools beyond Fleet’s 200+ built-ins). First wiki mention.
- Ramp + Prime Intellect — named on stage as the Qwen-3.5-fine-tuning collaboration; first wiki mentions.
- Pied Piper account — Caroline’s demo CRM example (named after the Silicon Valley HBO show); not promoted.
Concept pages touched:
- agent-harness — adds the three-layer continual-learning model (model / harness / context) as a structural sub-decomposition of the harness concept itself; adds the evals-as-gradient framing for harness-and-context-layer improvement; adds the third primary-source citation of the Top 30 → Top 5 LangChain Terminal Bench 2 claim.
- agent-development-lifecycle — extends the Build phase with the two-types-of-agents typology (long-horizon vs customer-experience); extends the Eval phase as the structural source of the gradient for continual learning on harness and context layers; the LangChain Labs announcement is the vendor commitment to continual learning as a first-class lifecycle phase.
- ai-agents — adds the two-types-of-agents structural typology (long-horizon vs customer-experience) — orthogonal to existing taxonomies; useful for product-design framing.
- enterprise-ai-adoption — adds the domain-experts-building-agents-not-just-giving-feedback thesis at the vendor-CEO altitude; the rebuilt-in-Fleet-so-the-domain-team-could-own-it migration is a worked example of the no-code-layer enabling org-design shift Hu and Tan articulate from the accelerator-CEO altitude.
- foundation-models — adds the rise of open-source models three-driver framing (capability / cost / trainability for particular domains); names the Fireworks AI partnership as the open-model commercial-deployment surface.
Source quality
- Channel: LangChain official YouTube — vendor channel; conference keynote format.
- Format: ~22-minute three-speaker keynote with a live demo; manual English captions (highest-quality transcript source in the wiki’s video-ingest pipeline).
- Empirical anchors: LangChain’s own GTM-agent metrics (84% weekly usage / 240% lead-to-qualified / 40h saved per rep per month) are vendor-cited and falsifiable only by LangChain. The Top 30 → Top 5 on Terminal Bench 2 claim is now anchored across three independent primary sources (Trivedy March 10, Chase Interrupt 26 stage assertion, and the cohort of harness > model corroborating papers).
- Bias / motive: Vendor-CEO keynote on the vendor’s own conference channel. Read every forward-looking claim (two-types typology, voice future, sandbox necessity, open-models rise, identity/auth patterns, continual learning) as a bet LangChain is making and product-strategising around — useful as forward-looking-intent data, not as neutral observation. The Fleet metrics are LangChain-internal and would benefit from a third-party comparison.
- Transcript provenance: youtube-transcript-skill (Playwright path); manual English captions automatically preferred over ASR; first ingest in this session where the Playwright path succeeded on first attempt without yt-dlp fallback.