[Harness] Agents enter unproductive loops (repeated tool calls, no progress)

[senamakel]

Summary

In large-scale aggressive fuzz testing with our internal agent-simulation harness, agents frequently enter unproductive loops: they re-issue the same (or trivially similar) tool calls repeatedly, make no progress toward the goal, and never terminate or change approach. The turn/step budget is consumed with no result.

This is a tracking issue for the "no-progress loop" failure class. Specific mechanisms are tracked in the sub-issues below.

Simulated example

A simulated agent asked to "find the service config and report the active port" loops on the same call:

→ list_dir(path="/app")
← [directory listing]
→ list_dir(path="/app")          # identical call, no new information
← [directory listing]
→ list_dir(path="/app")          # repeats ~15×, never opens a file
...
(step budget exhausted — no answer produced)

Impact

• Wasted compute and latency; step/turn budget exhausted with nothing returned.
• Appears "stuck" — no progress, no resolution, no message.

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: #4088 · #4089 · #4090