The original title is "How Monday.com Built Sidekick on Deep Agents | Interrupt 2026"

Short summary

Omri Bruchim (monday.com engineering manager) walks through how Sidekick evolved from a failing LangChain ReAct loop (200 tools, infinite context) to a production-grade Deep Agents system. They rebuilt using four principles: deferred tool discovery, delegation-first sub-agents, sandboxed code-writing, and self-healing mechanisms that achieved 94% error recovery. The talk covers real scaling failures and engineering patterns applicable to any multi-tool agentic system.

• •Sidekick V1 failed due to context pollution and tool explosion; rebuilt on Deep Agents architecture
• •Four core principles: deferred discovery, delegation-first, code-writing tools with LangSmith sandbox, self-healing with 94% recovery
• •Production-validated patterns for scaling multi-tool agents beyond toy systems

Generated with AI, which can make mistakes.