.@mondaydotcom built their AI copilot, Sidekick, to do more than answer question

monday.com, which makes work-management software, found that giving its AI assistant Sidekick more tools made it worse once it moved into production workflows. In a guest post on LangChain's blog, Omri Bruchim, who leads AI engineering at monday.com, says the first version was one general-purpose agent with a growing list of tools. In production, tools with overlapping descriptions made it harder for the model to pick the right one, tool definitions crowded out room for the user's request and data, one prompt had to cover every kind of task, and a failure in one step could derail a long workflow. The team rebuilt Sidekick by deciding, for each part of a request, which of three mechanisms should handle it.

Tools handle bounded actions with clear inputs and outputs, such as reading a board or updating an item; Sidekick's tools now sit in tiers that the agent must explicitly activate before it sees their full definitions. Specialized subagents, with their own instructions and smaller toolsets, handle narrower reasoning such as risk analysis or writing. Sandboxes, isolated workspaces from LangChain's LangSmith, handle work that produces intermediate files. If a user uploads CSV files to reconcile with board data, the agent can save them in a sandbox, write and run a script, fix a parsing error and generate a chart there, so the main agent receives only the result and a summary instead of every intermediate table. Bruchim says the design has run in production for a few months and describes the use cases it supports, but gives no figures on accuracy, cost or speed, and the post appears on the blog of the vendor whose components monday.com uses.