Phase 1: A complete, isolated environment stood up on demand that simulates the world your code will live in.
AI-generated software is verified in the environment it was built in. There's a gap between "tests pass" and "this actually works in a real-world scenario."
Agents claim things work because they have full context of what they just built. A real user starts from zero.
Issues go unsolved because agents aren't forced to consider deployment details — packaging, installs, configuration.
Integrating with Teams, Slack, or GSuite requires real accounts, app registrations, and rate limits. Unsustainable at scale for a software factory.
A complete, isolated environment stood up on demand from a declarative profile that simulates the world the code will live in.
"What will reality be like if this was actually deployed?"
mitmproxy-based redirection. GitHub URLs resolve to local Gitea repos transparently.
Build wheels from local source, serve via local pypiserver. pip/uv see your modified packages.
Forward real API keys (Anthropic, etc.) into the environment. No mocking needed for stateless services.
Setup commands to install tools, configure software, prepare the environment exactly as a real user would.
Mirror amplifier-core and amplifier-module-provider-anthropic from GitHub into an ephemeral Gitea instance.
Bump the amplifier-core version, add a print statement to the provider module. Changes live only in Gitea.
amplifier-digital-twin launch amplifier-user-sim --var GITEA_URL=... --var GITEA_TOKEN=...
uv tool install git+https://github.com/microsoft/amplifier — exactly as a real user would.
Run Amplifier. The modified amplifier-core wheel and rewritten provider are active — transparently.
It thinks it got everything from PyPI and GitHub. No special configuration. No test flags. Just reality, redirected.
Human or agent manages lifecycle: create / destroy
YAML declares what the universe contains
Isolated OS environment (Ubuntu 24.04)
Transparent URL rewriting
Serves locally-built wheels
Forwards real API credentials
Run commands or get an interactive shell inside the environment.
Monitor what's happening — logs, network, processes.
Drive user-like interactions from the outside.
All commands return JSON to stdout for programmatic consumption. Built for humans and agents alike.
Declarative YAML profiles define complete environments. Self-contained and launchable.
Redirect GitHub URLs to Gitea transparently. The software inside never knows.
wheel_from_git, wheel_path, wheel_var — multiple ways to serve modified packages.
Real external services (Anthropic, etc.) are accessible inside the twin with forwarded credentials.
Create, use, destroy on demand. Clean simulation every time.
Full end-to-end tests proving the complete flow works — from launch through install to verification.
uv tool install works correctly with your changes, exactly as a real user would experience it.Pre-built mock services for M365, Slack, GSuite. Unlimited use without rate limits, app registrations, or cost.
Graphical access — remote desktop or browser sessions. You can already exec into the container; this is about GUI-level interaction.
browser-tester, terminal-tester integration. Agents drive "as a user" experiences through the same surfaces as humans.
Full domain-level simulation. Mock services addressable by their real-world hostnames.