Context Engineering Our Way to Long-Horizon Agents: LangChain’s Harrison Chase

Watch on YouTube ↗  |  January 21, 2026 at 13:01  |  39:48  |  Sequoia Capital
Speakers
Harrison Chase — Cofounder, LangChain

Summary

Harrison Chase explains that long-horizon AI agents are finally working because better reasoning models are paired with improved harnesses and context engineering. He argues coding agents are the leading use case, with file systems, memory, traces/evals, and code sandboxes as core primitives. The conversation also covers how building agents differs from traditional software, the role of human judgment, and implications for existing software vendors and vertical AI startups.

  • Long-horizon agents are becoming practical as models and harnesses improve.
  • Coding is the most mature agent domain and is spreading to research, finance, and support.
  • Harness engineering, context compaction, file systems, and memory are critical agent primitives.
  • Traces and evals are becoming the source of truth for agent development.
  • Code sandboxes are preferred over browser-use agents for now.
  • Existing software companies with proprietary data and APIs can benefit from exposing data to agents.
  • Vertical AI agents with deep domain knowledge are seeing strong demand.
  • Long-horizon agents require async management plus sync chat and state viewing.
Ideas
Harrison Chase Cofounder, LangChain 2:06
Long-horizon agents finally work at scale
Harness engineering is becoming a critical layer: the same core LLM-in-a-loop algorithm works much better with opinionated planning, compaction, sub-agents, skills, MCP, and file-system tools, and harnesses often need to be tuned to model families like Anthropic or OpenAI, so third-party coding companies can create real performance gains.
AI-SECTOR HIGH
Harrison Chase Cofounder, LangChain 25:17
Data-rich incumbents gain from agents
Existing software companies with valuable proprietary data and good APIs should benefit because data is becoming more valuable and agents can plug into those APIs; however, the instructions for what to do with the data are a newer and harder part that may require new domain knowledge.
Data-rich software incumbents MED
Harrison Chase Cofounder, LangChain 34:37
Agent inbox needs sync and async
Long-horizon agents need both async management for many parallel runs and sync chat for corrections; viewing shared state like files is important, and LangChain's agent inbox with chat was a big unlock, so agent management UIs are an important emerging layer.
Agent inbox/agent management UI MED
Harrison Chase Cofounder, LangChain 38:31
Browser-use agents lag code execution
Browser-use agents are not yet good enough from what Harrison has seen; coding and code execution have worked much more than browser use in the short term, so he prefers code-execution-based approaches over browser-use agents for now.
Browser-use agents MED
Up Next
48:19
Quadrillions: How to Win the World | Chris Maurice
Empire · Jan 21
6:07:56
[26.01.21. 오후 방송 전체보기] 알테오젠, 기술이전 규모 및 로열티 기대치 하회에 오늘의 빌런 등극
3PRO TV (삼프로TV) · Jan 21
2:58
Davos: TCW’s Koch Sees ‘Marginal Move Away’ From US Treasuries
Bloomberg Markets · Jan 21
5:39
OpenAI CFO Sarah Friar: There's a massive 'capability overhang' in AI that's occurring
CNBC · Jan 21
3:05
Instability drives volatility as risk appetite fades and global markets reprice
CNBC · Jan 21

This Sequoia Capital video, published January 21, 2026, features Harrison Chase discussing AI-SECTOR, Data-rich software incumbents, Agent inbox/agent management UI, Browser-use agents. 4 trade ideas extracted by AI with direction and confidence scoring.

Speakers: Harrison Chase  · Tickers: AI-SECTOR, Data-rich software incumbents, Agent inbox/agent management UI, Browser-use agents