Sanity adds give_feedback tool to MCP server to collect agent bug reports
Sanity shipped a `give_feedback` MCP tool that lets AI agents report errors, missing capabilities, and confusing outputs directly to the team — no human required.
Score breakdown
The `give_feedback` pattern fills a blind spot in MCP server observability — telemetry can show which tools fail, but only an agent-callable feedback tool can surface whether the agent actually accomplished its goal or had to work around a gap.
- 01Sanity's MCP server has received over three million tool calls from more than 20,000 agents since going GA in December 2025.
- 02Existing qualitative feedback channels (Discord, GitHub, email surveys, docs) all assume a human is on the other end, leaving agent-generated issues unreported.
- 03The new `give_feedback` MCP tool was inspired by a post from Teddy Riker at Ramp on designing for agents.
Sanity's MCP server has become a significant platform for agentic work, with over 20,000 agents making three million tool calls since the server went GA in December 2025. Despite having telemetry on which tools get called and which fail, the team had no way to learn whether an agent actually accomplished its goal, where it got confused, or what capabilities were missing — and every existing qualitative feedback channel assumes a human is on the other end.
The solution was a `give_feedback` MCP tool, inspired by a post from Teddy Riker at Ramp on designing for agents.
The solution was a `give_feedback` MCP tool, inspired by a post from Teddy Riker at Ramp on designing for agents. The tool has one required field — `message`, capped at 2,000 characters — where the agent describes its goal, what it tried, and where it got stuck. Optional fields include `category` (with values `tool_error`, `missing_capability`, `confusing_output`, `documentation`, or `other`), `toolName` to identify the relevant tool, and `severity` (defaulting to `medium`, where `high` means the agent was blocked, `medium` means it found a workaround, and `low` is a minor annoyance). The agent receives a one-line confirmation and continues its task.
The post notes that shipping the tool alone is insufficient — agents must also be told when and how to use it. Sanity landed on three discovery channels: the tool listing itself, server instructions attached to the `initialize` response that starts every session, and error messages returned by other tools. The team reports the tool has already helped them improve the MCP server.
Key facts
- 01Sanity's MCP server has received over three million tool calls from more than 20,000 agents since going GA in December 2025.
- 02Existing qualitative feedback channels (Discord, GitHub, email surveys, docs) all assume a human is on the other end, leaving agent-generated issues unreported.
- 03The new `give_feedback` MCP tool was inspired by a post from Teddy Riker at Ramp on designing for agents.
- 04The tool has one required field: `message`, capped at 2,000 characters, describing the agent's goal, what it tried, and where it got stuck.
- 05Optional fields include `category` (`tool_error`, `missing_capability`, `confusing_output`, `documentation`, `other`), `toolName`, and `severity`.
- 06`severity` defaults to `medium`; `high` means the agent was blocked, `medium` means it found a workaround, and `low` is a minor annoyance.
- 07Agents discover the tool via the tool listing, server instructions in the `initialize` response, and error messages from other tools.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 19, 2026 · 10:25 UTC. How this works →