Harrison Chase
Confidence 0.85 · 3 sources · last confirmed 2026-06-20
Harrison Chase is the co-founder and CEO of LangChain — the company he started in 2022 around the open-source langchain Python framework, which expanded over 2023–26 into a layered product stack (LangChain framework / LangGraph runtime / Deep Agents harness pattern / LangSmith observability & evaluation / LangSmith Fleet no-code agent builder / LangChain Labs research group). The wiki’s canonical vendor-CEO voice on agent-engineering infrastructure — Chase coins the frameworks / runtimes / harnesses / no-code four-layer Build vocabulary and the model / harness / context three-layer continual-learning model that anchor the wiki’s agent-development-lifecycle and agent-harness concepts.
Promoted from Dangling to an entity page on 21 May 2026 after the second substantive source mention:
- [[2026-05-09-chase-agent-development-lifecycle|Chase / LangChain blog — The Agent Development Lifecycle (9 May 2026)]] — Chase’s foundational ADLC essay; named the wiki’s second formalization of the ADLC and the source most influential on the concept page’s current shape (per the LangChain entity page). Establishes the four-layer Build split (frameworks / runtimes / harnesses / no-code) that the wiki now uses as its working taxonomy.
- Chase, Sproul & di Vittorio — Interrupt 26 Day 2 keynote (21 May 2026) — 12 days later, Chase headlines the Interrupt 26 keynote with the two-types-of-agents typology (long-horizon vs customer-experience) and the three-layer continual-learning model (model / harness / context) — extending the ADLC vocabulary along two new structural axes. Announces LangChain Labs as a research group inside LangChain “aimed in particular at continual learning.”
Role in the wiki
Chase operates at the vendor-CEO-of-agent-engineering-infrastructure altitude — articulating the operating vocabulary for what is being built while shipping the product stack that instantiates each layer. Three distinct contributions surface across the two ingested sources:
1. The four-layer Build vocabulary (ADLC, 9 May)
Chase’s foundational structural decomposition of the Build phase of the agent lifecycle:
| Layer | What it is | LangChain product |
|---|---|---|
| Frameworks | Abstractions for model calls, tools, prompts, retrieval, structured outputs, agent loops | LangChain (the framework) |
| Runtimes | Stateful agent runtime; control flow / branching / looping / persistence / pause-resume | LangGraph |
| Harnesses | Prompts, skills, MCP servers, hooks, middleware, memory, virtual filesystem | Deep Agents |
| No-code | Domain-expert-facing UI to create and configure agents without code | LangSmith Fleet |
This taxonomy is now the wiki’s canonical Build-phase vocabulary on agent-development-lifecycle, with LangChain serving as the vendor whose product taxonomy is the wiki’s vocabulary — both because Chase coined the four-layer split and because LangChain ships at all four layers.
2. The two-types-of-agents typology (Interrupt 26, 21 May)
A parallel-vertical typology orthogonal to the four-layer Build split: Chase predicts (and observes already emerging) a divergence in agent product surfaces:
- Long-horizon agents — run for minutes / hours / days; code execution, planning, sub-agents, multi-agent, skills; outcomes and goals as horizon-extension mechanisms; knowledge work.
- Customer-experience agents — latency-critical; voice (and maybe video) as modality; brand-sensitive; sales / customer support / 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?“
3. The three-layer continual-learning model (Interrupt 26, 21 May)
The Interrupt 26 keynote’s conceptual centerpiece: every agentic system has three independently improvable layers, each with its own continual-learning loop:
| Layer | What it is | Improvement mechanism |
|---|---|---|
| Model | The base foundation model (Sonnet, GLM4, GPT-4) | Fine-tuning on traces (Chase cites Ramp + Prime Intellect fine-tuning Qwen 3.5 for Ramp sheets as the worked example) |
| Harness | ”The code surrounding the model that connects it to the environment” (Deep Agents, Claude Code, pi) | Agent-edits-harness with evals as forcing function — Chase cites MetaHarness (MIT + Stanford) as the canonical paper |
| Context | ”What we provide to the harness as ways to guide it on particular tasks” (agent.md, skills) | Agent.md / skill files iterated against eval suites |
The evals-as-gradient framing (Chase’s own articulation): “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 evals and traces are incredibly important for this learning.”
4. The on-stage confirmation of the LangChain Top 30 → Top 5 claim
Chase asserts at Interrupt 26: “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.” This is the third primary-source confirmation of the same numerical claim — first asserted in LangChain Engineering blog (10 March 2026), and the strongest single-claim convergence in the wiki’s harness > model corpus.
Notable framings Chase has contributed
- “Frameworks / Runtimes / Harnesses / No-code” — the four-layer Build vocabulary now used as the wiki’s working taxonomy.
- “Model / Harness / Context” — the three-layer continual-learning model.
- “Evals act as a forcing function” — the gradient analogy for non-model-layer learning.
- “Two types of agents: long-horizon and customer experience” — the product-surface typology orthogonal to the build-layer split.
- “Everyone will be involved in building agents — domain experts most of all” — the org-side corollary of layered-agent thinking, paired with Diana Hu’s AI founder / IC / DRI org structure prescription.
Career timeline
- Pre-LangChain — “My background’s in classical machine learning” (Chase, Interrupt 26, ~9:20).
- 2022 — Founded LangChain as an open-source Python framework for working with LLMs.
- 2023–2024 — LangGraph (stateful agent runtime) and LangSmith (observability + eval) launched as the product stack grows out of the open-source base.
- 2025–2026 — Deep Agents pattern, LangSmith Fleet (no-code), LangChain Academy, LangSmith Sandboxes shipped. Company organises Interrupt 2026 (this source) as its first major industry conference.
- 2026 — Announces LangChain Labs as a research group inside LangChain aimed in particular at continual learning (Interrupt 26).
Convergence with wiki sources
| Source | Connection |
|---|---|
| ADLC (9 May 2026) | The foundational four-layer Build vocabulary; wiki’s primary Chase source for the ADLC concept page. |
| Interrupt 26 (21 May 2026) | The continual-learning extension; two-types-of-agents typology; LangChain Labs announcement; vendor-CEO confirmation of Top 30 → Top 5. |
| LangChain Engineering (10 March 2026) | LangChain colleague’s first articulation of the Agent = Model + Harness coinage that Chase’s Interrupt 26 three-layer model extends. |
| LangChain Engineering — Improving Deep Agents with Harness Engineering (17 Feb 2026) | The deep-agents-as-harness pattern that became the substrate for LangSmith Fleet (per Brace Sproul, Interrupt 26 ~17:23). |
| MetaHarness (30 March 2026) | The academic paper Chase cites by name on stage at Interrupt 26 as the canonical learn at the harness layer worked example. |
| Continual Harness (11 May 2026) | The academic formalisation of continual-harness adaptation — Chase’s Interrupt 26 three-layer model is the vendor-CEO operationalisation 10 days later. |
| Stanford CS153 (20 May 2026) | Paired founder/CEO articulations of layered-agent-system thinking landing within 24 hours. Tan/Hu maps agentic-primitives → company-structure; Chase maps learning-systems → agentic-system-layers. Same architectural decompose-and-map move applied to different domains. |
Open questions
- Chase’s pre-LangChain career — “My background’s in classical machine learning” is a one-line autobiographical aside; the wiki has no further detail on where Chase worked before founding LangChain in 2022.
- LangChain Labs’s first publications — Labs is announced at Interrupt 26 as the continual-learning research group inside LangChain; the first Labs outputs (papers, blog posts, open-source releases) are an open ingest target for late 2026 / 2027.
- LangSmith Fleet customer adoption — the internal-LangChain GTM-agent metrics (84% weekly usage / 240% lead-to-qualified / 40h saved per rep / month) are the only quantified Fleet outcomes the wiki holds. Customer-side adoption metrics (any LangChain customers using Fleet at scale with published numbers) would substantiate or refine the no-code-builder-for-domain-experts thesis.
- LangSmith Engine — referenced in the Interrupt 26 description as a separate Interrupt-week announcement (“Meet LangSmith Engine” blog link); not substantively discussed in the talk body. Open ingest target.