Internal Agents Map

Design Agent

Amplitude built Design Agent, an internal web tool that anyone in the company reaches behind Google OAuth to turn a text prompt, a screenshot, or a rough product idea into on-brand HTML output. Amplitude's brand guidelines and design-system tokens sit in the agent system prompt, so the output reads as the company's own. The tool began as a proof of concept for product and engineering teams and has expanded to other designers, marketers, and creative teams.

Company
Amplitude
Approach type
Agent
Work
Design
Human involvement
Drafts reviewed
Invocation
Interactive
Interfaces
Web
Deployment stage
Deployed
Evidence strength
Limited primary
Entry reviewed
Copy as MDAsk ChatGPT

How it works

Prompt or screenshot in the internal web app, through managed-agent generation and R2 storage, to an artifact a person shares or carries forward.

Submit the request in the web interface

The interface offers a text input, an option to upload a screenshot or reference image, and a panel that displays the generated artifact.[1]

Let the managed agent plan and generate

Claude Managed Agents reasons about the prompt, runs a multi-step plan, handles tool-call follow-ups, and assembles the final artifact.[1]

Persist the artifact with a stable URL

Every generated artifact lands in Cloudflare R2 with a stable, permanent URL, and users keep a visible generation history.[1]

Share, iterate, or carry the output forward

People share the output links in Slack threads, return to generate follow-ups on previous outputs, or copy the generated HTML into a coding agent.[1]

Where people stay involved

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

Reported: A person writes every request and then judges the artifact, sharing the link, re-prompting, or pasting the HTML into a coding agent; the post documents no separate approval gate.

  • Supervision level

    Work-product review

    A person reviews the draft or implementation.

    • prompt or screenshot → generated HTML artifact a person judges, shares, or carries forward[1]

How supervision levels work

Implementation details

Model
Claude, through Claude Managed Agents; no model version is named.[1]
Harness
Claude Managed Agents supplies reasoning, tool use, and multi-step generation; Amplitude wrote no state machine, prompt chain, or tool-calling logic. A thin wrapper of Cloudflare Workers serves the web interface and the agent interaction endpoint.[1]
Sandbox
Not reportedThe post names Claude Managed Agents as the runtime but no execution isolation boundary.
Tool access
Amplitude states that it defined the agent's tools and context but does not name them. Generated artifacts are written to Cloudflare R2, each with a permanent URL.[1]
Knowledge
A structured Markdown document in the Google design.md format holds Amplitude's brand guidelines, color system, typography rules, spacing conventions, and component patterns.[1]
Context management
The brand and design-system document is baked into the agent system prompt, so it is present in every conversation; behavior changes ship by editing that prompt and take effect in minutes.[1]
Credentials
Not reportedGoogle OAuth gates human access to the web app; how the agent itself authenticates to Claude Managed Agents, Workers, or R2 is not documented.
Interfaces
  • web
[1]

Mechanisms

Prompt or screenshot to on-brand HTML artifact

A text prompt, a screenshot, or a rough product idea becomes interactive HTML styled to the Amplitude design system.[1]

Bake brand context into the system prompt

Amplitude's design philosophy and design-system tokens are written into the system prompt, which Amplitude reports turned generic model output into output that reads as its own.[1]

Validation and failure handling

Instrument every session and turn failures into regression tests

Amplitude tracks every session, output, and tool call with its own analytics, reviews sessions to find failure patterns, and creates eval examples from the worst outputs as regression tests before updating skills files, prompts, and agent configuration.[1]

Reported observations

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

Headline claim

Over 2,219 session snapshots in Design Agent's first few weeks, with repeat usage across multiple teams (self-reported)[1]

Reported by
Amplitude
Scope
Design Agent session snapshots across Amplitude teams in the first few weeks of use, as reported in the May 2026 post
Denominator
None reported; the post gives no user, team, or session total to divide by
Method
Amplitude product analytics instrumented on Design Agent, which the post says tracks every session, output, and tool call

Observed in May 2026

About this result
Category
Adoption output
Evidence type
Reported measurement
What it covers
Design Agent session snapshots across Amplitude teams in the first few weeks of use
Key observation

Amplitude reports roughly 2-4x more viewers than makers for generated artifacts, which it reads as a signal that outputs are shared beyond the people who generate them[1]

Reported by
Amplitude
Scope
Viewers of generated Design Agent artifacts compared with the people who generated them
Denominator
None reported
Method
Not stated; the post reports the ratio without a counting rule

Observed in May 2026

About this result
Category
Adoption output
Evidence type
Estimate
What it covers
Ratio of viewers of generated artifacts to the people who generated them
Key observation

Amplitude reports a working Design Agent for its product and engineering teams two days after starting, on Claude Managed Agents and Cloudflare[1]

Reported by
Amplitude
Scope
Elapsed time from starting the build to a working Design Agent for Amplitude product and engineering teams
Denominator
Not applicable
Method
Not stated; the author reports the duration without a record of hours or a defined start

Observed in May 2026

About this result
Category
Implementation scale
Evidence type
Qualitative
What it covers
Elapsed build time to a working Design Agent for product and engineering teams

Lessons

Amplitude wrote its brand guidelines and design-system tokens into the agent system prompt and reports that the difference from generic model output was drastic; this is its own result for one internal design tool, not a measured comparison.[1]

Amplitude instruments Design Agent with its own product analytics, reviews sessions to find failure patterns, and turns the worst outputs into eval examples that act as regression tests.[1]

Amplitude reports that managed agent orchestration and serverless hosting left it with no servers, monitoring, autoscaling, or database to run, which suits an internal tool maintained by one person.[1]

Reported opinion: Will Newton concludes that agents are products needing repeated iteration with real users, and advises keeping the stack simple and not skipping instrumentation; this is his stated view from building this one internal tool.[1]

Back to agents
Improve this record

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

Suggest on GitHub

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. How we built a design agent at Amplitude with Claude managed agents and Cloudflarehttps://www.amplitude.com/blog/design-agentEngineering blog · First party · Last source verification: 2026-09-21
    Preserved copy
Question coverage and scope
purpose
Reported
workflow
Reported
human involvement
Reported: A person writes every request and then judges the artifact, sharing the link, re-prompting, or pasting the HTML into a coding agent; the post documents no separate approval gate.
implementation
Reported
validation
Reported
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
    Amplitude reports the count from its own analytics instrumentation on Design Agent, with no independent review and no total of eligible users, teams, or sessions.
    Reported by
    Amplitude
    Scope
    Design Agent session snapshots across Amplitude teams in the first few weeks of use, as reported in the May 2026 post
    Denominator
    None reported; the post gives no user, team, or session total to divide by
    Method
    Amplitude product analytics instrumented on Design Agent, which the post says tracks every session, output, and tool call
    Observation date
    2026-05
  3. Harness
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  4. Model
    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. Context management
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  9. Prompt or screenshot to on-brand HTML artifact
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  10. Submit the request in the web interface
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  11. Let the managed agent plan and generate
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  12. Persist the artifact with a stable URL
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  13. Share, iterate, or carry the output forward
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  14. Bake brand context into the system prompt
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Not assessed
    Confidence reason
    No explicit confidence assessment is recorded for this claim.
  15. Instrument every session and turn failures into regression tests
    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
    Low
    Confidence reason
    The post gives the ratio as roughly 2-4x with no period, no counting rule for a viewer or a maker, and no underlying totals, so the range is an approximation rather than a stated measurement.
    Reported by
    Amplitude
    Scope
    Viewers of generated Design Agent artifacts compared with the people who generated them
    Denominator
    None reported
    Method
    Not stated; the post reports the ratio without a counting rule
    Observation date
    2026-05
  17. Key observation
    Statement type
    Metric
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    The author states the two-day figure twice as elapsed build time for the first working version; it describes implementation effort rather than a performance result, and no start date or effort total is given.
    Reported by
    Amplitude
    Scope
    Elapsed time from starting the build to a working Design Agent for Amplitude product and engineering teams
    Denominator
    Not applicable
    Method
    Not stated; the author reports the duration without a record of hours or a defined start
    Observation date
    2026-05
  18. Lesson
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    The post describes the design.md context document and the system-prompt change as steps the team took, and calls the difference drastic; that judgment is the author's own, with no side-by-side evaluation.
  19. Lesson
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    The measurement section states the practice directly: sessions reviewed in Agent Analytics for failure patterns, eval examples created from the worst outputs as regression tests, then skills files, prompts, and configuration updated.
  20. Lesson
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    The hosting section states the absence of servers, monitoring, autoscaling, and a database and ties it to a one-person internal tool; this is the team's own description of its setup, not an audited operating record.
  21. Lesson
    Statement type
    Opinion
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    The closing section is the author's stated conclusion from this one build; he offers no comparison with other agents, so the advice is attributed rather than established.
  22. Operating model assessment
    Statement type
    Inference
    Provenance
    Catalog judgment
    Confidence
    Medium
    Confidence reason
    The source shows the agent producing a finished artifact that the requester then judges, shares in Slack, re-prompts, or copies into a coding agent; it documents no separate approval gate, so the boundary is read from the described use rather than from a stated review step.
    Observation date
    2026-05