Uploaded June 2026 | Updated November 2026, 1 day ago
Omri Bruchim, engineering manager at monday.com, walks through the full journey of building Sidekick — monday.com's intelligent personal assistant — from a simple LangChain React loop all the way to a production-grade agentic system powered by Deep Agents. He covers the real scaling failures that forced a rebuild, and the four engineering principles that got them to a 94% error recovery rate in production.
Chapters:
0:00 Introduction and what we're covering today
0:31 About Omri and the monday.com AI platform
1:25 monday.com's mission shift: from managing work to doing the work
2:02 The SDR agent and job recruiter agent demos
2:22 The monday.com product suite: Monday Agent, AI Apps, and Sidekick
2:39 What Sidekick actually does
3:06 Sidekick V1: the LangChain ReAct loop that started it all
3:54 Why V1 broke: 200 tools, infinite context, context pollution
4:37 Why they chose Deep Agents to rebuild from scratch
6:55 Core philosophy: one smart orchestrator, not a swarm
7:41 The four principles — overview
7:44 Principle 1: Deferred tool discovery (the three-tier system)
9:55 Principle 2: Delegation first — sub-agents and async handoff
11:03 Delegation first — the middleware pipeline
12:04 Principle 3: The code-writing tool and LangSmith sandbox
13:04 Real-world example: finding a London office with live code
13:42 Principle 4: Self-healing with 94% recovery success rate
15:07 What's next for Sidekick
15:26 Takeaways for building with Deep Agents
Resources:
→ Deep Agents: langchain.com/deep-agents
→ LangSmith: langchain.com/langsmith
→ LangChain Academy: academy.langchain.com
→ monday.com AI platform: monday.com/ai
Omri Bruchim, engineering manager at monday.com, walks through the full journey of building Sidekick — monday.com's intelligent personal assistant — from a simple LangChain React loop all the way to a production-grade agentic system powered by Deep Agents. He covers the real scaling failures that forced a rebuild, and the four engineering principles that got them to a 94% error recovery rate in production.
Chapters:
0:00 Introduction and what we're covering today
0:31 About Omri and the monday.com AI platform
1:25 monday.com's mission shift: from managing work to doing the work
2:02 The SDR agent and job recruiter agent demos
2:22 The monday.com product suite: Monday Agent, AI Apps, and Sidekick
2:39 What Sidekick actually does
3:06 Sidekick V1: the LangChain ReAct loop that started it all
3:54 Why V1 broke: 200 tools, infinite context, context pollution
4:37 Why they chose Deep Agents to rebuild from scratch
6:55 Core philosophy: one smart orchestrator, not a swarm
7:41 The four principles — overview
7:44 Principle 1: Deferred tool discovery (the three-tier system)
9:55 Principle 2: Delegation first — sub-agents and async handoff
11:03 Delegation first — the middleware pipeline
12:04 Principle 3: The code-writing tool and LangSmith sandbox
13:04 Real-world example: finding a London office with live code
13:42 Principle 4: Self-healing with 94% recovery success rate
15:07 What's next for Sidekick
15:26 Takeaways for building with Deep Agents
Resources:
→ Deep Agents: langchain.com/deep-agents
→ LangSmith: langchain.com/langsmith
→ LangChain Academy: academy.langchain.com
→ monday.com AI platform: monday.com/ai










