New blog post! Our platform team’s story of rearchitecting the MCP Server to get past the wall that every agentic system eventually hits: Dynamic Agents: Addressing the Tool-Call Ceiling with MCP Code Mode
Pre-built tool chaining looks great in the early demos. Then the record sets get bigger, the questions get harder, and the context window starts eating your agent alive. Familiar?
This post lays out how we collapsed our growing catalog of MCP tools down to three, handed the LLM a sandboxed runtime, and let it compose its own logic against platform primitives at runtime. Give it a read—we’d love your take in the comments.