← 回總覽

基于 Deep Agents 和 LangChain 构建的社区 GTM 智能体

📅 2026-03-31 01:16 Harrison Chase 人工智能 3 分鐘 2847 字 評分: 82
LangChain AI 智能体 GTM Deep Agents 开源
📌 一句话摘要 Harrison Chase 重点介绍了社区利用 LangChain、Deepline 和 Deep Agents 构建的开源 GTM 智能体,展示了实用的智能体工作流。 📝 详细摘要 Harrison Chase 分享了一个由社区构建的“市场进入”(GTM)智能体,该智能体利用了 LangChain、Deepline 和 Deep Agents。引用的推文详细拆解了其技术架构,包括 Slack API 集成、对话记忆和提供商延迟等方面的挑战。这为开发智能体工作流的开发者提供了一个实用的案例研究,并附带了开源仓库。 📊 文章信息 AI 评分:82 来源:Harrison

Title: Community-Built GTM Agent Using Deep Agents and LangChain...

URL Source: https://www.bestblogs.dev/status/2038666570772459997

Published Time: 2026-03-30 17:16:01

Markdown Content: I rebuilt @LangChain GTM agent on @DeeplineCLI + Deep Agents in under an hour. ... Then spent 2 hours debugging the Slack API. But such is life.

Comment repo or DM me if you want the setup guide & configs I didn't include in the public repo.

Here's what worked, what broke, and what I'd do differently.

LangChain published how their internal GTM agent drove 250% more lead conversions. blog.langchain.com/how-we-built-l…

The architecture is solid. Deep Agents for orchestration, bunch of data sources for context, Slack for rep approvals.

If you wanted to build this from scratch - majority of the time would be figuring out what data sources + providers to use, testing those, and then customizing the prompts for your use cases.

That's what Deepline does for agents like Claude Code & Deep Agents. One API key, tool calls for 32+ providers. Ex. Waterfall enrichment, CRM, outreach tools.

So I wired Deep Agents to Deepline's API, and had a working waterfall enrichment agent on @Railway in < 30min (also first time using that, huge fan).

~200 lines of config.

What actually went wrong:

  • Slackbot - Spent an hour just getting the signing secret + bot token wired correctly with event & interactivity hooks. Would just use a manifest template if I did this again.
  • Conversation memory. Will add something like Redis. Without it, every Slack thread isn't very stateful.
  • Provider timing, less so rate limiting. Some enrichment tasks take 8-10 seconds to respond. The waterfall is 10 providers deep. If you're unlucky and the first 3 timeout, a simple "find the email for Jane at Ramp" takes 30+ seconds. Need better task management.
The system prompt is doing a lot of heavy lifting. When the agent picks the wrong tool, it's usually because the prompt didn't disambiguate well enough. Still tuning this, as Claude Code was doing a lot of heavy lifting for the CLI + Skills

What still needs work:

  • Batch operations. Right now it's one contact at a time. Need to add CSV upload + async processing.
  • Better error messages. When a provider fails, the agent just says "I couldn't find that." Not helpful. Should say which providers it tried and why they failed.
  • Cost tracking. No visibility into how much each query costs across providers.
Open sourced it: github.com/getaero-io/dee…

If you're building GTM agents and want something that just works, might save you some time!

Any agentic GTM workflows you're trying to solve right now?

查看原文 → 發佈: 2026-03-31 01:16:01 收錄: 2026-03-31 04:00:14

🤖 問 AI

針對這篇文章提問,AI 會根據文章內容回答。按 Ctrl+Enter 送出。