Agents in the Wild

This directory is no longer updated. Last update: November 16, 2026.

Duolingo's Slack agent grew from about 20 users to more than 250 a week, about 30% of the company

Duolingo put its internal tools behind one AI app in Slack. It answers in helpdesk and incident channels. Any write action outside Slack needs a person's approval. Duolingo open-sourced the core code.

Organization
Duolingo
Industry
Education technology
When
Sep 2025
Evidence
Company reported

Reported outcome

More than 250 weekly active users (about 30% of the company) by April 2026, up from around 20 in the first week. The upvote rate on answers stabilized around 80%.

  • more than 250 (about 30% of the company)
    Weekly active users (April 2026)
  • around 20
    Users in the first week (September 2025)
  • around 80%
    Upvote rate on responses
  • around 20 cases
    Regression test suite

Caveats

Stated by the company that uses the agent (blog, earnings call, interview). Not independently audited.

  • The numbers come from a talk by a Duolingo engineer, written up by the Agentic AI Foundation. Nobody outside checked them.
  • The write-up does not agree with itself on tool counts. The title says 180+ MCP tools. The body says 30 MCP servers with 300+ tools, and that the app has access to 15+ MCP servers.
  • The upvote rate counts only people who pressed the feedback button. It is not an accuracy rate.
  • The write-up says the bot cut toil for on-call engineers but gives no number for it.

What the agent does

Duolingo engineers wanted to use MCP servers, which connect AI tools to company systems. Setting each one up by hand was slow and often failed. Two earlier fixes, a discovery page (May 2025) and standard hosting for every server (August 2025), did not move adoption.

In September 2025 the team built an AI Slack app and put it in the channels people already used. It runs on the Claude Agent SDK and the Slack Bot SDK, and it can query BigQuery for internal metrics. It answers by itself in helpdesk and incident channels. Each team can adjust how it behaves in its own channel.

Why this works

  • Nobody has to configure anything. The tools come to the place where people already work.
  • Any write action outside Slack (a PR, a ticket, a deploy to staging) needs a person's approval in a confirmation screen.
  • The bot does not get around existing permissions. Access follows job roles, and it runs in a sandboxed VM.
  • Every answer has a feedback button. A regression suite of around 20 cases runs all the time and grows.

What to copy

  • Remove the setup step. Put the agent where the work already happens.
  • Let the agent read freely within a person's permissions, and ask before it writes.
  • Add a feedback button to every answer and turn bad answers into test cases.

Starter prompt

Written by us, not by Duolingo. Connect only tools the asking user is already allowed to use.

You answer questions in the {channel} Slack channel.
Tools you can use: {tools}

Answer from tool results only. Name the tool behind each fact.
If a request needs a write outside Slack (PR, ticket, deploy), describe the exact action and ask for approval. Do not run it.
If you cannot find the answer, say so and name the team to ask.

Sources

  1. How Duolingo Built an AI Slackbot With 180+ MCP Tools — Agentic AI Foundation, Apr 27, 2026

Added Nov 16, 2026 · Sources last checked Nov 16, 2026.

Get new use cases weekly

One short email. Sources included.

The newsletter is paused. Past issues stay in the archive.

Related use cases