AI agents at Stripe that transform Slack messages into code changes, shipping roughly 1,300 pull requests per week from team conversations to production.
A Slack reaction became an agent trigger: Stripe engineers could summon AI minions that pick up bounded engineering tasks and turn them into reviewable pull requests.
The signature pattern is a low-friction trigger paired with serious delivery controls. An engineer signals intent in Slack, the system extracts or routes the task, and an agent executes against repository context until it has a pull request worth reviewing.
The magic is not that Slack can start an agent. It is that the interface lowers task capture to almost zero while the output still lands in the normal engineering quality system: pull requests, tests, and human review. That is how the workflow can scale without turning into a side channel.
Use Slack as the intake surface so engineers do not leave the conversation to create an AI task.
const task = {
trigger: "slack_reaction",
threadUrl,
requester,
repository,
acceptanceCriteria
};The workflow should reject vague or high-risk requests and send only reviewable tasks to the coding agent.
A successful minion run creates a PR with tests, summary, and reviewer context instead of silently changing production behavior.
| Tool | Version | Role | Why this tool |
|---|---|---|---|
| Slack | Current | Intake | Makes starting an AI task as cheap as reacting in the conversation where work is discussed. |
| Coding agents | Current | Execution | Transforms small tickets into concrete repo diffs. |
| GitHub pull requests | Current | Review surface | Preserves ownership, traceability, and existing engineering controls. |
| CI checks | Current | Validation | Keeps agent output accountable to the same gates as human-authored code. |
SystemYou convert a bounded engineering request into a small pull request.
User templateFrom this Slack thread, infer the task, inspect the repo, implement the smallest fix, run checks, and prepare a PR summary.
Privacy. Public discussion on the How I AI podcast; specifics drawn from public episode content.