A failed coding-agent experiment exposed problems in our MCP design

Описание и применение

Эксперимент с кодирующим агентом выявил проблемы интерфейса MCP в FuncHole: неясные описания инструментов, отсутствие защитных механизмов для shell‑команд, нехватка примеров структуры функций и сложности с статическими/фронтенд‑функциями.

Задача

Designing an interface for coding agents to build, deploy, test, and iterate applications via FuncHole was insufficient, leading to agent failures.

Результат

The experiment highlighted that exposing tools to an agent is not enough; the interface must provide clear contracts, examples, constraints, guardrails, and actionable errors to support successful agent operation.

Рабочий процесс

• 1. Provide an end‑user requirement to the agent via FuncHole.
• 2. Agent attempts to use FuncHole’s MCP tools to build, deploy, test, and iterate the application.
• 3. Agent encounters unclear tool descriptions and falls back to shell commands without guardrails.
• 4. Agent struggles with missing examples of function structure and static/frontend functions.
• 5. Agent eventually deploys a working application, exposing interface shortcomings.

Доказательства

The agent struggled with several parts of our MCP interface: Some tool descriptions weren't clear enough. It tried falling back to shell commands where we lacked proper guardrails. We hadn't provided good examples of the expected Function structure. Static/frontend Functions were particularly difficult to reason about.

Источник