06-reference

technically what does langchain do

2026-10-13·reference·source: Technically·by David Krevitt
langchainlangsmithagent-harnessobservabilityai-tooling

Technically — What does LangChain do?

A standalone explainer (not part of Technically's Railway-sponsored "Software Eng for Vibe Coders" series) by David Krevitt — the publication footer reads "© 2026 Justin" (Justin Gage, the outlet's legal/masthead signature), but the byline and voice match Krevitt's prior Vibe Coders Parts 2-4 bylines. No sponsor block, no "brought to you by" language, and no ad placement anywhere in the rendered HTML (checked beyond the plaintext email, which only carries boilerplate CSS referencing a sponsorship-campaign-embed class with no sponsor content behind it) — consistent with Technically's unsponsored-by-default posture outside the Vibe Coders series, matching the same-format precedents [[2026-09-29-technically-whats-duckdb]] and [[2026-05-21-technically-what-is-nextjs]].

Fidelity note: the free email/web preview cuts off mid-article ("Continue reading this post for free in the Substack app... Or upgrade your subscription") after the observability section. The business-model, funding, and Databricks/dbt-competitive-positioning discussion previewed in the TL;DR is behind that paywall and not captured here — this note covers the free portion only (TL;DR, the pre-LangChain chatbot/RAG pain points, and the LangSmith observability pivot).

Why this is in the vault

LangChain/LangSmith is the dominant commercial attempt to productize exactly the layer RDCO builds ad hoc for itself — agent harness, tracing, and eval tooling — so this explainer is a direct read on how a well-funded vendor frames the problem RDCO's own sub-agent fan-out and /improve loop solve informally.

The core argument

Krevitt frames LangChain as "basically everything but the model itself," now spanning three layers: (1) open-source frameworks (LangChain, LangGraph) for building LLM apps; (2) observability/eval tooling (LangSmith) to monitor and score agents; (3) infrastructure (sandboxes, a model router, deployment tooling) to run agents in production. He positions LangChain as following Databricks and dbt down the "long, tough road" of building a company around an open-source project — a comparison the piece doesn't develop further before the paywall.

The history: early LLM API products (2023, GPT-3-era) were simple chatbots wired to RAG pipelines, but four problems kept surfacing — RAG tuning (chunking strategy), prompt save/reuse/versioning, guardrails (models wouldn't reliably follow format instructions), and model-switching (OpenAI vs. Anthropic). LangChain's original framework solved for those. But solving that layer immediately exposed the next one: LLMs are non-deterministic, so teams needed observability to see what agent sessions were actually doing. That's what LangSmith started as — tracing input through output — and its newer "Engine" product closes the loop by having an LLM read accumulated traces and suggest improvements to the app itself. Krevitt names Comet and PostHog's self-driving mode as competing players chasing the same "self-improving LLM app" observability category.

Mapping against Ray Data Co

Concrete gap exposed: RDCO has no LangSmith-equivalent trace/eval layer for its own sub-agent fan-outs, and this piece names exactly why that matters. RDCO doesn't run LangChain/LangGraph — the harness is Claude Code + the Claude Agent SDK directly (cf. [[2026-04-06-claude-agent-sdk-guide]]) — but the problem LangSmith was built to solve (can't tell what a non-deterministic agent actually did across a session, can't systematically score whether its output was good) is unsolved for RDCO's own fan-out pattern, including this very skill. The "Workflow agent output integrity" memory already names the symptom from the inside (pointer returns, scratchpad contamination, false "verified" stamps) and mitigates it with a hand-rolled rule — "one gate per chain MUST hit the primary source" — rather than a trace/eval system. LangSmith's Engine (trace accumulation → LLM-suggested app improvements) is a more structured version of what /improve tries to do by reading feedback and diarizing mediocre responses; worth a look as a model for tightening that loop rather than as a tool to adopt.

The Databricks/dbt open-source-monetization comparison is flagged but not assessable from the free preview — it's the kind of thing worth a follow-up read if the question of how agent-tooling vendors monetize (relevant to any future phData positioning on the "who owns the harness layer" question) comes up again.

Related