[Harness] Agents don't escalate / hand back control when stuck

[senamakel]

Summary

In fuzz testing with our internal agent-simulation harness, agents that get stuck rarely escalate or hand back control. Instead of surfacing the blocker to the orchestrator/Core or to the user, they keep retrying silently until the budget runs out.

This is a tracking issue for the missing-escalation failure class; mechanisms are in the sub-issues.

Simulated example

→ attempt_action()  → blocked
→ attempt_action()  → blocked
→ attempt_action()  → blocked
(agent never says "I'm blocked on X — handing back" and never asks for input)

Impact

• Failures are invisible: no handoff, no question to the user, no fallback path taken.
• The run dead-ends instead of recovering via escalation.

Sub-issues

Tracked below.

---

Surfaced by our internal agent-simulation harness during large-scale, aggressive fuzz testing of agent behaviors. The example above is synthetic and contains no real data.

Sub-issues: #4091 · #4092