Investigate augmenting coding agents for ADK development

[tomkis]

Is your feature request related to a problem? Please describe.
Working on ADK with coding agents (e.g. Claude Code) requires too much manual intervention. The agent can't self-sufficiently run, debug, and validate changes against the full stack. Developers still need to manually orchestrate dev environments — starting servers, connecting UI, setting up telepresence — before the agent can be productive.

Describe the solution you'd like
Two key areas to investigate:

1. Better feedback loops for agents — Agents should be able to autonomously run ADK, use all existing debugging capabilities (telepresence, e2e tests, integration tests), and validate their own changes without human hand-holding. Today we have adk-server-debugging skill and worktree skill but agents still can't close the full build→run→test→debug loop independently.

2. Automated dev environment preparation — A developer should be able to say "prepare me the env for UI development with server changes" and the agent handles everything: starting the VM, connecting telepresence, running the server, connecting the UI in dev mode, etc. No manual steps. This applies to any dev scenario (server-only, UI+server, full stack, etc.).

Describe alternatives you've considered

• Continuing to manually set up environments and guide agents through debugging steps
• Writing more detailed skill instructions (helps but doesn't close the autonomy gap)

Additional context
Existing infrastructure to build on:

• mise run adk-server:dev:start/connect/disconnect/reconnect/reload tasks
• mise run adk-ui:run:dev for UI development
• .claude/skills/adk-server-debugging and .claude/skills/worktree skills
• Telepresence-based debugging with Lima/MicroShift VMs
• E2E and integration test suites