MCPJam Engineer: 'MCP Is Bad at X' Conflates Three Very Different Failures
Paoli99 · reddit · 2026-10-09
An MCPJam engineer building agent evals argues that 'MCP vs CLI' debates lump together distinct problems: passing data between tools (structured output helps, but the next tool must accept it), context crowding from tool definitions and results (their evals don't yet measure this), and picking the right tool. For tool selection they test both that the agent calls the expected tool and avoids the wrong one, flip test cases, and run repeats to check consistency. Key caveat: single-step evals passing doesn't mean multi-step chaining works, nor that code orchestration wouldn't be better.
More from coding & agent
- OpenAI-Affiliated Account Teases Major Agent Progress: "A Helluva 3 Months" Ahead — iruletheworldmo · 2026-10-09
- Agent Plasticity paper: frontier models differ wildly in how efficiently they self-improve from experience — _akhaliq · 2026-10-09
- Rebuild any e-commerce cart with Claude + Aftersell's copy-prompt workflow — EXM7777 · 2026-10-09
- Running agents across machines: give the agent SSH keys to other machines it can use at will — lucasmeijer · 2026-10-09
- Open-source Android app Angel runs MCP servers on-device and hands tools to local or cloud models — ihaveaboyfriendsorry · 2026-10-09
- ClickUp's Brain² agent builds reports and dashboards with E2B microVM sandboxes — mathemagic1an · 2026-10-09