Improve architecture and state coupling between web-ag-ui and agents

[0xTomDaniel]

Summary

Improve architecture and reduce state/coupling inconsistencies between web-ag-ui and agents.

Problems observed

• Frontend appeared to manage 6 runs at once while only 3 agents were present.
• Sidebar status polling should run about every 15 seconds, but behavior looked duplicated/unbounded.
• agent-long-graph-agent.connect appeared to keep a persisted streaming connection even when the agent details page was not in focus.
• Streaming connection should only remain open while the specific agent details page is open; it should close on navigation away.
• Agent-side state transitions feel inconsistent/chaotic across agents.

Goals

• Define and enforce clear state lifecycle rules on both frontend and agent sides.
• Reduce duplicated polling and run-management fan-out in the UI.
• Ensure streaming connections are opened/closed strictly by page focus and lifecycle.
• Identify shared agent state-management patterns and factor common constraints into reusable architecture.
• Refactor UI architecture where needed to reduce coupling and improve maintainability.

Acceptance criteria

• Polling cadence and ownership are centralized and deterministic.
• Number of managed runs aligns with actual active agents/runs.
• Streaming connections reliably tear down when the details page is not active.
• Agent state transitions follow a documented/common pattern with reduced divergence.
• Architecture/documentation updates include concrete boundaries between UI orchestration and agent orchestration.