Tool Design & MCP Integration

Two book-search tools, both with one-line descriptions; the agent picks between them at 60% accuracy.

A payment MCP returns "Operation failed" on every error — and the agent's recovery behavior is erratic.

One powerful general-purpose web tool, or three narrow purpose-built ones?

Configuring MCP for a shared Jira, a personal Notion workspace, and an experimental scratch server.

A new MCP code-analysis tool exists, but the agent keeps reaching for built-in Grep instead.

A 22-tool finance agent picks the right one only 64% of the time — even with good descriptions.

An empty array can mean "no orders" or "system was down" — the agent treats them identically and gets it wrong.

Every transient timeout in a subagent is propagated all the way up — and latency is suffering.

Every conversation begins with the agent loading 200 client records it might not need.

Six invoice-processing tools, one strict ordering rule, and a 4% skip rate on the prerequisite step.

An Edit fails on non-unique anchor text and the agent gives up. What's the documented fallback?

An agent reaches for `Bash find` to enumerate test files in a 5,000-file codebase — and gets node_modules back.

An MCP tool returns `{success: false}` and the agent occasionally treats it as a successful response.

A synthesis subagent inherits ten tools — and decides to email the synthesis output to the user.

There's a maintained community Jira MCP server. A teammate proposes building a custom one "for full control."