LangChain on why company-level agent harnesses are hard, and how Managed Deep Agents fixes it
hwchase17 · x · 2026-10-21
LangChain founder HW Chase highlights a post from colleague Caspar on why building a "company harness" is hard, alongside the launch of Managed Deep Agents (MDA).
Drawing on internal agents deployed across LangChain's marketing, sales, and engineering teams, Caspar lays out the core primitives every multiplayer agent harness needs: living where people already work (Slack), acting on-behalf-of the requester for some services and with its own identity for others, memory scoped globally or per-user, sandboxed code execution where authenticated CLIs (gh, curl, internal APIs) just work, and a large share of runs triggered on schedules.
MDA is LangChain's managed runtime: developers define the agent as a directory of files — agent.py, instructions.md, skills, tools, middleware — while LangSmith Agent Server runs the deep agent harness (planning, filesystem, subagent delegation), threads, runs, streaming, and an MCP endpoint with no servers to operate.
More from coding & agent
- Kronos: open-source K-line foundation model trained on 12B candles, with a free AI trading setup guide — Roger_M_Taylor · 2026-10-21
- $33.49 Claude-made hype video remake landed a $450k deal — full open-source pipeline — zealcaiden · 2026-10-21
- Dev builds 5 projects in a day without hitting Opus 5.5 or GPT 6.1 usage caps — joshgonsalves_ · 2026-10-21
- ego browser claims 47% faster agent web automation at $1.64 per task — kimmonismus · 2026-10-21
- Claude Code autonomously researches Cursor, Windsurf and Zed into clean tables — kimmonismus · 2026-10-21
- ego (lite) beta brings agent-driven browsing to Claude Code and Codex on Windows — kimmonismus · 2026-10-21