Loop Control And Pivots

Loop Control and Pivots

Repeating a failing approach burns budget and hides progress. Fail fast, pivot on evidence, and stay on the goal.

Activation triggers

• The same approach has failed ~2–3 times (same error, same dead end).
• You are grinding a side problem — env setup, a missing tool, credentials, a file transfer, working around tooling — instead of the actual deliverable.
• A debug / exploit-dev / build / research thread is not converging.

The rules

1. 3-strikes → dead: after ~3 evidence-based attempts at one approach, mark the path dead. Do not re-trigger it blindly; quote the failure and switch to a different test, not a repeat. New evidence or partial progress resets the count — strikes track repeated identical failure, not a path that is still yielding signal.
2. Fail fast on non-retryable failures: some errors never resolve by repetition — auth/permission denials, 404/not-found, malformed-input or unsupported-feature rejections. Spend zero retries: fix the cause, pivot, or surface [BLOCKED: need X] at once instead of burning the 3-strikes budget.
3. Pivot on evidence: failed path → quote the exact failure → next shortest path. Each retry must change a variable, not just re-run hope.
4. Don't grind side quests: a secondary problem gets one bounded attempt (see Budgets). If it doesn't yield, surface [BLOCKED: need X] and pivot to productive work rather than sinking the run into it.
5. Hold the objective: every step ties to the success signal. If you've drifted onto a sub-problem, name the drift and return to the goal (or escalate for the missing capability).
6. Escalate honestly: when genuinely stuck after the recovery pass, pivots, and local tests, load the narrowest hint/research support skill or hand back an honest blocker with everything derived so far (offsets, leaks, partial output) — never a fabricated success.

Persist on the objective (the other half of fail-fast)

Fail-fast applies to a path, never to the objective. A dead path means pivot; it does not license abandoning the goal. The dominant failure mode on capable models against hard targets is premature surrender — bailing with "stuck" / "need a tool" / BLOCKED while budget and untried attack-classes remain.

• Two levels, opposite reflexes. Pivot the vector/class fast (3-strikes, above); keep attacking the objective until the run budget is committed or every distinct attack-class has been tried and killed with evidence. Do not collapse "this path is dead" into "I am stuck."
• Kill paths with tests, not proofs. "Dead" requires a failing live experiment. A reasoning-based "this can't work / is unreachable / is impossible" verdict is a hypothesis, not a kill — abandoning the correct vector on a deductive impossibility is a top cause of losing solvable targets. When your attempts have become impossibility arguments instead of experiments, the pivot is to dynamic probing (hook the candidate call site, fuzz the threshold/length/format), not deeper static reasoning. (See hypothesis-driven → "'Impossible' is a hypothesis".)
• Recovery pass before BLOCKED. On a solvable target, before you report blocked: (a) diff current state vs your hypothesis and probe what is still unverified; (b) re-read recon for missed env vars, comments, headers, versions; (c) try the simplest attack of the class, not the cleverest; (d) send wild/empirical payloads — errors leak parser and structure.
• What BLOCKED actually means. A genuine blocker is a missing external capability — credential, authorization, access, or a tool you cannot obtain — not "out of ideas." Out-of-ideas → recovery pass or a hint/research skill; missing-capability → the BLOCKED report shape below.

Budgets and hard stops

Set caps before starting. Trip a cap → stop, quote it, pivot or report.

• Tool-calls per approach: ~5–10 calls to prove or kill one hypothesis. Past that: path is dead.
• Side-quest cap: ~10 tool-calls or ~15 min on env/tooling/creds/transfer. Past that: [BLOCKED: need X].
• Wall-clock / token budget: when ~70% of the run budget is spent, drop low-yield branches and reserve remainder for verification and reporting.
• MCP / network retry cap: at most 2 automatic retries on the same call with the same args. The third attempt must change the call, target, or route — otherwise mark it dead.
• Sub-agent runaway: kill and reassign a delegated worker when it exceeds its budget, produces no new artifacts across two status reports, or drifts off the stated goal.

Anti-patterns

Blocked report shape

[BLOCKED: need X] — what you were trying, the exact failure (verbatim), what you derived, and the one capability/decision that would unblock it. Pair with evidence-before-claims so the blocker is as auditable as a success would be.