The Loop Reliability Model
The Loop Reliability Model judges any agent loop on five axes: completion rate (does work finish), verification quality (is done actually done), recoverability (what happens on failure), observability (can you see what happened), and boundedness (can it run away). A loop weak on any axis fails in that axis's characteristic way.
Key takeaways
- The five axes are diagnostic: name the weak axis and you've named the failure mode.
- Verification quality dominates: a loop that completes confidently and wrongly is worse than one that escalates honestly.
- Boundedness is the governance axis: budgets, iteration caps, and abort criteria.
- Observability is what makes the other four checkable: unlogged loops can't be evaluated or debugged.
- Every Agentiquette loop-pattern page carries an assessment against this model.
The model
Any agent loop can be judged on five axes. Each axis has a characteristic failure, which is what makes the model diagnostic rather than merely evaluative: name the weak axis and you have named the failure mode you'll see in production.
| Axis | Question | Characteristic failure when weak |
|---|---|---|
| Completion rate | Does work actually finish? | Abandoned tasks, endless wandering |
| Verification quality | Is "done" actually done? | Confident false success |
| Recoverability | What happens on failure? | One flake kills the run, or silent skips |
| Observability | Can you see what happened? | Undebuggable, unevaluable outcomes |
| Boundedness | Can it run away? | Burned budgets, runaway side effects |
Reading the axes
Verification quality dominates. A loop that completes 9 of 10 tasks and knows which one failed is production infrastructure. A loop that claims 10 of 10 with one silent failure is a liability wearing a success metric. This is why the axis is defined by evidence: pre-stated criteria evaluated on captured output, per the verification step discipline. All other axes limit damage; this one prevents it.
Recoverability is a spectrum, not a boolean. Retry-with-diagnosis, backtrack, escalate, abort: each is legitimate, and the failure is having none of them. The cheapest upgrade to most loops is a mandatory diagnosis step between attempt and retry, which is the entire insight of budgeted retry.
Observability is the enabling axis. Unlogged loops cannot be debugged, cannot be benchmarked, and cannot support the audit story governance needs. The standard: every outcome in the log should be explainable from the log.
Boundedness is governance inside the loop. Iteration caps, cost budgets, abort criteria, and an escalation path when they trip. Persistence is the point of a loop; unbounded persistence is an incident.
Applying it
To assess an existing loop, ask the five questions against evidence from real runs, not against the loop's documentation. To compare loops, hold the task set fixed and measure: completion rate, false-success rate, recovery rate under injected failures, and log completeness. That measurement design is exactly what Lab benchmarks implement.
As a worked example, plan-execute-verify assessed against the model: strong on verification quality (criteria precede execution) and boundedness (step budgets are native), moderate on recoverability (diagnosis-then-retry, but backtracking needs explicit design), strong on observability (the plan-versus-actual record is the audit trail), and its completion rate is the empirical question our first benchmark measures.
Every loop-pattern page in the Patterns Library carries an assessment against this model. When you spec your own loop, the loop spec template forces answers on all five axes before the loop ever runs.
Frequently asked questions
What is loop reliability?
The quality of an agent loop measured on five axes: completion rate, verification quality, recoverability, observability, boundedness.
Which axis matters most?
Verification quality, because its failure mode (confident false success) is the most expensive. The other axes limit damage; verification prevents it.
How do you measure a loop's reliability?
Run a fixed task set through it and measure completion rate, false-success rate, recovery rate under injected failures, and whether the logs explain every outcome. This is exactly what Agentiquette Lab benchmarks do.
Summary
- Topic
- The Loop Reliability Model
- Definition
- Five axes for loop quality: completion, verification, recoverability, observability, boundedness
- Best used for
- Comparing loop patterns and diagnosing loop failures
- Related concepts
- agent loop, plan-execute-verify, false success
- Common mistakes
- Optimizing completion while ignoring verification; unbounded loops; no logs
- Recommendation
- Assess your loop on all five axes; fix verification first, boundedness second