Agent Stack: a shared space to test and improve agents
Made sharing an agent as easy as pasting a URL
Overview
- Role
- Design Engineer · product strategy, UX, prototyping
- Period
- 2025 Q4
- Tools
- Figma · React · TypeScript · Next.js
What I did
- Set the product direction around one loop: share → try → respond
- Designed the lifecycle across validation, build, configuration, live use, recovery, and ownership
- Built the product experience with streaming builds, secret injection, and live agent chat
Problem
Agents lived as repositories and container images that only their authors could run. Sharing meant setup instructions, CLI steps, secrets managed by hand, and no shared place to try what existed. The platform had to work across models and frameworks while making sharing as simple as pasting a URL.
Users
Agent builders
Researchers and developers need to put working agents in front of teammates without writing setup instructions or choosing one model stack.
Teammates and reviewers
Colleagues need a zero-setup way to try the behavior, inspect its evidence, and respond while the agent is still changing.
What I designed & built
I designed Agent Stack as a collaboration product, not a deployment dashboard. It validates and builds a portable artifact, asks for configuration only when needed, then opens the live agent directly into a surface teammates can use and respond to.
Interaction architecture
One happy path, every consequential branch
The happy path and recovery states were designed together, covering unreachable sources, failed builds, missing secrets, updates, and removal.
Add source
Repo or image URL
Validate
Reachability first
Build
Stream real work
Configure
Only what is missing
Try + respond
Live team feedback
↳ Recovery
Invalid URLs, private repos, build failures, and missing configuration each resolve to a specific next action.
↳ Ownership
Existing agents expose source, environment variables, update paths, and destructive actions in one place.
↳ Consumer experience
Published agents open as usable products, with cited answers and a lightweight response-feedback loop.
Process
Share the source, not setup instructions
A builder shares one repository or container image. Agent Stack validates reachability before starting the build, keeping the entry point model-agnostic and surfacing private or missing sources immediately.
Turn infrastructure into a product state
Instead of a spinner or fake estimate, the interface streams actual build events. The same trace explains failures, and every wait resolves into a named state with a clear next action.
Ask for configuration at the moment of need
The built artifact declares its missing keys only after a successful build. Save and go live injects encrypted values, reruns the health check, and publishes the agent in one action.
Close the loop from response to review
Teammates can inspect a cited response, score its helpfulness, choose a category, and leave a natural-language note. Builders see the results in an agent-specific feedback table and can download the complete review set as a CSV.
Learnings
- Model-agnostic UX starts with portable artifacts and lets each agent declare its own configuration.
- Streaming real work and naming every state makes uncertain build time understandable.
- The real finish line is teammate feedback, not deployment.
Next steps
- Test the share → try → feedback loop with active agent teams and measure time to first teammate response.
- Design versioning and rollback so builders can iterate without breaking the agent teammates are evaluating.
- Connect build events and secret discovery to the real platform API, then pressure-test recovery with failed and partial builds.