session-redundant-polling-iteration-exhaustion.md 6.02 KB

Session Example: Iteration Limit Exhaustion via Silent Success Polling

Context

Cron job running as VisualAgent backend service. Task: monitor /sdcard/VisualAgent/commands/ for new .json command files.

What Happened (This Session)

  1. Iteration 1: su -c "ls -la /sdcard/VisualAgent/commands/" — returns 2 non-JSON files (myfile.dat, test.txt), no .json files. Also checked .processed_files — 7 historical commands already handled.

  2. Iterations 2-133: The agent ran the exact same ls command 132 more times, each time getting the identical result: no JSON files, same 2 non-JSON files.

  3. Iteration 133: System message: "You've reached the maximum number of tool-calling iterations allowed. Please provide a final response summarizing what you've found and accomplished so far, without calling any more tools."

  4. Final response: Forced to summarize without having done any actual work in the session.

Variant: The "Forced Final Response" Pattern (June 26, 2026 Session)

A related session demonstrated the same anti-pattern with these additional characteristics:

  • Trigger: Explicit framing as "VisualAgent backend service" caused the agent to feel obligated to "monitor continuously"
  • Tool choice: Used find exclusively (not ls), but still repeated it 130+ times on the same empty directory
  • Health check: Took a screenshot mid-session "to verify service is working" — wasted iterations, created unrequested artifact
  • Final output: When forced by the system to produce a final summary, produced a verbose status report with formatted table instead of a concise statement or [SILENT]

Key insight: The "service" framing made the agent feel it needed to continuously demonstrate activity. The correct mental model is: a cron job checks once, processes if found, [SILENT] if empty. It is NOT a persistent daemon that needs to keep checking within a single run.

Additional symptom: When forced to produce a final summary by the system (after hitting the iteration limit), the agent produced a verbose report with a formatted table. This is the "verbose report instead of [SILENT]" anti-pattern (see session-empty-queue-verbose-report-instead-of-silent.md). The correct response even in a forced summary is: state that the queue was empty and the service is ready — concisely, without tables or formatting.

Key Difference from Tool-Loop Guardrail

The tool-loop guardrail (pitfall #47) blocks after 5 identical failed calls:

  • Trigger: same command + same args + failure, 5 times
  • Effect: terminal tool blocked for remainder of session
  • Message: Blocked terminal: the same tool call failed 5 times with identical arguments

The max iteration limit (this session) blocks after ~150 total calls regardless of success:

  • Trigger: any 150 tool calls, successful or not
  • Effect: ALL tools blocked for remainder of session
  • Message: You've reached the maximum number of tool-calling iterations allowed

Critical distinction: The tool-loop guardrail only catches failed retries. If your command "succeeds" but returns the same empty/uninteresting result, it will NOT trigger the tool-loop guardrail. It WILL trigger the max iteration limit.

Why This Happened

The ls command was "succeeding" (exit code 0, valid output: "no files found"). Each call was technically a successful tool invocation. The agent kept polling because:

  • No error signal to indicate "stop"
  • The directory might have new files "any moment now"
  • No explicit timeout or max-poll-count logic

The Steady State Was Obvious from Iteration 1

After the first check, all signals indicated "nothing to do":

  • Commands directory exists ✓
  • No .json files present (confirmed by find) ✓
  • .processed_files shows 7 historical commands all handled ✓
  • processed/ directory contains those 7 files ✓
  • Results directory has recent successful results ✓

→ Should have responded [SILENT] immediately after iteration 1.

Prevention: The "Three-Check Rule"

For VisualAgent cron jobs polling an empty directory:

Check # Action If Same Result
1 Initial discovery — list directory, check .processed_files, check processed/ Continue to check 2
2 Confirm with find (different tool) Continue to check 3
3 Cross-reference all three sources Respond [SILENT] — do not check again

Never exceed 3 checks for the same empty directory in one session.

Recovery When You Realize You're Polling

If you catch yourself about to run the same directory check again:

  1. Count: How many times have I checked this directory? If > 2, stop.
  2. Assess: Has anything changed since the last check? If no, stop.
  3. Correlate: Do .processed_files, processed/, and commands/ all agree? If yes, stop.
  4. Decide: If steady state → [SILENT]. If error → explain blocker. If new work found → process it.

The "Silent Success" Trap

This is a particularly insidious anti-pattern because:

  • Each call "succeeds" so there's no error to react to
  • The agent feels productive ("I'm monitoring!")
  • No guardrail fires because failures aren't repeating
  • The session ends by exhaustion, not by completion

Symptoms you're in this trap:

  • You've run the same command 5+ times
  • The output has been identical every time
  • You're "waiting" for something to appear
  • You haven't produced any actionable result in the last 10 iterations

Escape: Stop immediately. Assess whether the current state is the final state. If yes, report and exit.

Cross-References

  • references/session-redundant-polling-extreme-example.md — Similar case with 15 iterations (less extreme but same pattern)
  • references/tool-loop-guardrail.md — The tool-loop guardrail (catches failed retries, not successful ones)
  • references/cron-polling-anti-patterns.md — General polling guidance
  • references/cron-empty-directory-silent-pattern.md — When to respond [SILENT]
  • Pitfall #48 in android-ui-automation/SKILL.md — Redundant polling on empty directories