Internal Agents Map

Minions

Homegrown one-shot coding agents that read work context, implement a change end-to-end, and produce a PR for human review.

Company
Stripe
Approach type
Agent
Human involvement
Drafts reviewed
Interfaces
Slack, Github, Cli, Web, Internal ui
Deployment stage
Scaled
Evidence strength
Detailed primary
Entry reviewed
Copy as MDAsk ChatGPT

How it works

An engineer sends a request through Slack or an internal app. The agent prepares a pull request for human review.

Start a run from Slack or an internal app

An engineer tags the Slack app from the thread discussing a change, or starts a minion from the CLI, the web interface, or an integrated internal tool such as the flaky-test ticketing UI[Source 1]

Hydrate context before the run

Relevant MCP tools run deterministically over likely-looking links before the run starts, and rule files scoped to subdirectories attach as the agent traverses the filesystem[Source 1, Source 2]

Implement the task in a prewarmed devbox

The blueprint's agent nodes implement the change in an isolated devbox, interleaved with deterministic lint, git, and push nodes[Source 2]

Open the pull request for human review

The minion prepares a pull request following Stripe's PR template; the engineer opens it and requests review, or sends further instructions that push updated code[Source 1]

Where people stay involved

People review, approve, or guide the agent’s work here.

Reported: Runs are unattended; the engineer opens the prepared pull request for review or returns it with further instructions.

  • Supervision level

    Work-product review

    A person reviews the draft or implementation.

    Level 3 of 5

    • coding task → pull request handed to an engineer for review[Source 1]

How supervision levels work

Implementation details

Model
Not reported
Harness
Fork of Block's goose, orchestrated by code-defined blueprints that interleave agent loops with deterministic lint, git, and CI steps[Source 2]
Sandbox
Pre-warmed AWS EC2 devboxes in the QA environment, isolated from real user data, production services, and arbitrary network egress[Source 2]
Tool access
Curated subsets of Toolshed MCP tools for internal documentation, tickets, build status, and code intelligence; security controls constrain destructive actions[Source 2]
Knowledge
Repository-scoped rule files shared with human-operated coding agents, plus internal context fetched through MCP[Source 2]
Context management
Subdirectory-scoped rule files attach as the agent traverses the filesystem, relevant MCP tools run over links before the run starts, and blueprint nodes can constrain tools or simplify each subagent's context[Source 1, Source 2]
Credentials
Full permissions inside quarantined devboxes; MCP security controls limit destructive actions, and production pull requests require human review[Source 2]
Interfaces

Slack, CLI, web, and internal tools are named; the pull-request host is implied by the branch-and-PR flow but never named in the captures.

  • slack
  • github
  • cli
  • web
  • internal-ui
[Source 1]

Mechanisms

One-shot end-to-end agent

Prepares a branch for human review and returns after a limited number of CI repair attempts[Source 1, Source 2]

Validation and failure handling

Repair CI failures within a bounded loop

After a push, CI applies autofixes and sends unfixable failures back once for local repair; after at most two CI rounds the branch returns to its human operator[Source 1, Source 2]

Reported observations

These results come from the sources. Notes below each result explain what it covers and any reported limits.

Headline claim

Over 1,300 completely minion-produced PRs merged per week in Part 2, with human review and no human-written code[Source 2]

Reported by
Stripe
Scope
Weekly merged PRs completely produced by Minions; Part 2 report, up from Part 1

The source does not report the denominator of this figure.

About this result
Category
Adoption output
Evidence type
Reported measurement
What it covers
Weekly merged minion-produced PRs at Stripe, human-reviewed, per the Part 2 report
Key observation

Over 1,000 completely minion-produced PRs merged per week in Part 1, with human review and no human-written code[Source 1]

Reported by
Stripe
Scope
Weekly merged PRs completely produced by Minions; earlier Part 1 report

The source does not report the denominator of this figure.

About this result
Category
Adoption output
Evidence type
Reported measurement
What it covers
Weekly merged minion-produced PRs at Stripe, per the earlier Part 1 report

Lessons

Stripe's Minions produce changes unattended, then return them for engineer review; failed runs can receive further instructions.[Source 1]

Stripe starts Minions in prewarmed devboxes with recent copies of its main repositories and prepared development services.[Source 2]

Back to agents
Improve this record

Give your coding agent the context and instructions to prepare a source-backed update.

Paste the prompt into your coding agent, then describe the correction or add a public source.

Preview prompt

Sources and research details

Citations link to the original publisher. Each source also keeps a preserved copy in the repository, so a changed or removed page stays checkable.

  1. Minions: Stripe's one-shot end-to-end coding agentshttps://stripe.dev/blog/minions-stripes-one-shot-end-to-end-coding-agentsEngineering blog · First party · Last source verification: 2026-08-31
  2. Minions - Part Twohttps://stripe.dev/blog/minions-stripes-one-shot-end-to-end-coding-agents-part-2Engineering blog · First party · Last source verification: 2026-08-31
  3. Hacker News discussion of Minions part twohttps://news.ycombinator.com/item?id=47086557Hn thread · Community · Last source verification: 2026-08-31
  4. Hacker News comment asking for examples and review evidencehttps://news.ycombinator.com/item?id=47086907Hn comment · Community · Last source verification: 2026-08-31
  5. Hacker News comment questioning the level of technical detailhttps://news.ycombinator.com/item?id=47086953Hn comment · Community · Last source verification: 2026-08-31
  6. Hacker News comment asking about shipped and maintained outputhttps://news.ycombinator.com/item?id=47087114Hn comment · Community · Last source verification: 2026-08-31
  7. Hacker News comment questioning the value of pull request countshttps://news.ycombinator.com/item?id=47111453Hn comment · Community · Last source verification: 2026-08-31
Question coverage and scope
purpose
Reported
workflow
Reported: The one-shot characterization and the bounded CI repair loop sit under mechanism and validation; the four steps above are the run in reading order.
human involvement
Reported: Runs are unattended; the engineer opens the prepared pull request for review or returns it with further instructions.
implementation
Reported
validation
Reported: A deterministic local lint node and one or two CI rounds with autofixes check the work before the branch returns to its human operator.
observations
Reported
lessons
Reported
Research details for every claim on this page
  1. Summary
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  2. Headline claim
    Statement type
    Metric
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    Stripe reports merged PR volume without an independent count or measurement window; Part 2 is explicitly later than Part 1.
    Reported by
    Stripe
    Scope
    Weekly merged PRs completely produced by Minions; Part 2 report, up from Part 1
    Denominator
    Not reported
    Method
    Not reported
    Observation date
    Not reported
  3. Sandbox
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  4. Harness
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  5. Interfaces
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  6. Tool access
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  7. Knowledge
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  8. Credentials
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  9. Context management
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    Part 1 states the pre-run MCP hydration and Part 2 states the subdirectory-scoped rule files and blueprint-level context control directly.
  10. One-shot end-to-end agent
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  11. Start a run from Slack or an internal app
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  12. Hydrate context before the run
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  13. Implement the task in a prewarmed devbox
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  14. Repair CI failures within a bounded loop
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  15. Open the pull request for human review
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  16. Key observation
    Statement type
    Metric
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    Stripe reports merged PR volume without an independent count or measurement window; no calendar metric date appears in the capture.
    Reported by
    Stripe
    Scope
    Weekly merged PRs completely produced by Minions; earlier Part 1 report
    Denominator
    Not reported
    Method
    Not reported
    Observation date
    Not reported
  17. Lesson
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    Part 1 distinguishes unattended implementation from human PR review and explicitly describes another instruction cycle when the result is insufficient.
  18. Lesson
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    Part 2 describes proactive repository cloning and environment setup; it does not say the agent chooses the repository or provisions its own workspace.
  19. Operating model assessment
    Statement type
    Inference
    Provenance
    Catalog judgment
    Confidence
    High
    Confidence reason
    The source explicitly states that humans review and approve production pull requests.
    Observation date
    2026-02-20