Loop Breaker: Stop Repeating, Start Solving

Agents waste enormous effort looping through the same failing approach -- retrying identical strategies, seeing the same error, going in circles 10, 20, even 58 times. This skill breaks the cycle.

The Iron Law

The third identical attempt is always wrong. If you've tried something twice and it failed, doing it a third time is not persistence -- it is insanity.

The Process

After every failed attempt, check: have I tried this before?

digraph loop_breaker {
    rankdir=TB
    node [shape=box style=rounded]

    attempt [label="Attempt fails"]
    check [label="Have I tried this\napproach before?" shape=diamond]
    first [label="Log approach +\nerror in tried list.\nRetry ONCE with\ntargeted adjustment."]
    second [label="STOP. 2 failures\nwith same approach.\nTrigger pivot protocol."]
    list [label="1. List all tried approaches\n2. State WHY each failed\n3. Generate 3 fundamentally\n   different strategies\n4. Pick most promising\n   untried approach"]
    pivots [label="Have I pivoted\n3 times with\nno progress?" shape=diamond]
    escalate [label="STOP. Escalate\nto the user.\nPresent findings."]
    execute [label="Execute new\napproach"]

    attempt -> check
    check -> first [label="No / once"]
    check -> second [label="Yes, twice+"]
    second -> list
    first -> attempt [label="fails again" style=dashed]
    list -> pivots
    pivots -> escalate [label="Yes"]
    pivots -> execute [label="No"]
    execute -> attempt [label="fails" style=dashed]
}

Step 1: Track what you have tried

Maintain a mental ledger after each failure:

• Approach: What you did (specific file, function, strategy)
• Result: The exact error or failure mode
• Root cause: WHY it failed, not just THAT it failed

Step 2: Detect the loop (2-strike rule)

If the same approach has failed twice (same file, same strategy, same class of error), you are in a loop. Stop immediately. Do not attempt a third time.

Step 3: Force a pivot

When the 2-strike rule triggers:

1. List every approach tried and its failure reason
2. Diagnose the common thread -- why does this category of fix keep failing?
3. Generate 3 fundamentally different strategies -- not variations, genuinely different angles (different files, different tools, different assumptions)
4. Select the most promising untried strategy and execute it

Step 4: Escalate after 3 failed pivots

If you have pivoted to 3 genuinely different strategies and none produced progress, stop and escalate to the user. Present:

• What was tried (all approaches)
• Why each failed
• What you suspect the root cause is
• What information or access you need to proceed

Red Flags: Rationalizations That Signal a Loop

What Counts as "the Same Approach"

Two attempts are the same approach if they share any of these:

• Editing the same lines in the same file with minor variations
• Applying the same category of fix (e.g., adding imports, adjusting types)
• Getting the same error message or same class of error
• Operating on the same assumption about root cause

A fundamentally different approach means changing your assumption about what is wrong, not just how you fix it.