Reference
Glossary
60 terms | 6 clusters
Plain-language definitions for the AI agent reliability lane: failure modes, error propagation, evaluation rigor, and the terms behind the research.
Multi-agent failures
15 terms- Consensus and voting reliability Consensus and voting reliability is the cascade-resistance score applied to a voting ensemble: how far a majority lowers a fault's odds of reaching the output. The Condorcet result holds only when ballots are independent, and a majority only counts its ballots and never audits them.
- Context handoff loss (agent to agent) Context handoff loss is context that existed on the sending side of an agent-to-agent handoff and is missing from the brief the receiver acts on, so the receiver proceeds on a smaller brief than the sender held, with no error raised.
- Conversation reset (agent sessions) Conversation reset is the multi-agent failure mode in which a dialogue restarts before its task is finished and the state built inside it is discarded, so the agents resume from an earlier position and the progress made since is lost.
- Failure to ask for clarification Failure to ask for clarification is the multi-agent failure mode in which an agent meets an ambiguous brief, settles the ambiguity silently with a guess, and passes the guess downstream as a fact, so no agent after it learns that a question was ever open.
- Ignored agent input Ignored agent input is the multi-agent failure mode in which one agent contributes a correct and relevant signal, the receiving agent has it in full, and the system proceeds as though it had never been sent.
- Information withholding (multi-agent) Information withholding is the multi-agent failure mode where an agent holding context a downstream agent needs leaves it out of the handoff, so the receiver acts on an incomplete brief without ever learning the gap exists.
- Inter-agent misalignment Inter-agent misalignment is the multi-agent failure category in which coordination between agents breaks down: context a peer needed is withheld, correct input is ignored, an agent acts against the plan it just stated, and the shared objective drifts hop by hop.
- Multi-agent debate failure Multi-agent debate failure is the mode where correlated agents debating a task converge on a confident wrong answer, so the exchange ratifies a shared error and returns it as consensus. The result is hallucinated consensus: agreement a transcript cannot distinguish from a correct result.
- Orchestrator-worker reliability Orchestrator-worker reliability is how well a fan-out (hub-and-spoke) topology keeps one worker's faulty return from reaching its siblings. Its worst-case propagation radius approaches the fan-out width when the orchestrator redistributes an unvalidated return as a subtask input.
- Premature termination (AI agents) Premature termination in an AI agent system is a run that stops while the task's objectives are still unmet and returns the partial state as a completed result, so the shortfall never surfaces as an error.
- Reasoning-action mismatch Reasoning-action mismatch is the multi-agent failure mode in which an agent's stated plan and its executed action diverge, so the reasoning in the trace reads as correct while the step the agent actually took was a different step.
- Role drift (multi-agent systems) Role drift is the gradual erosion of an agent's assigned role across a long run, where every turn still returns well-formed output while the remit the agent is actually working to widens or slides away from the brief it was given.
- Specification violation (agent tasks) Specification violation is an agent failure in which the returned work breaks a constraint the task stated, such as an output format, a length bound, a forbidden tool or a scope limit, scored against the instruction rather than against whether the answer is correct.
- Task derailment Task derailment is the multi-agent failure mode in which the objective itself moves: each hop nudges the goal slightly, and after several hops the system is competently solving a problem adjacent to the one it was given, with no single step that reads as wrong.
- Task verification failure (agents) Task verification failure is the failure mode where an agent system's checking step returns a pass on output that does not satisfy the task, so a wrong result travels downstream carrying a confirmation and every agent after it treats the work as already checked.
Propagation & containment
15 terms- Backpressure (agent pipelines) Backpressure is the signal a saturated stage in an agent pipeline sends back to whatever feeds it, asking the producer to lower its rate so the queue between the two stays bounded and the pipeline refuses work at its entrance rather than losing it mid-run.
- Bulkhead isolation (agent systems) Bulkhead isolation is the practice of partitioning the resources an agent system shares, including worker slots, provider rate budgets, tool credentials and state, so that one agent's failure exhausts only its own partition while the rest of the system keeps running.
- Cascade resistance Cascade resistance is a topology-level resilience score for how well an architecture keeps one fault from becoming a system failure, ranking chain, supervisor, orchestrator-worker, and debate designs on one axis.
- Circuit breaker (AI agents) A circuit breaker in an AI agent system is a control that stops the orchestrator from calling a tool or sub-agent once that dependency's failures cross a stated threshold, holds the calls off through a cooldown, then admits a few probe calls before resuming.
- Containment rate Containment rate is the fraction of an injected fault a system stops before it crosses the first hop, reported with a confidence interval rather than as a bare point estimate.
- Error budget (agent systems) An error budget for an agent system is the count of task failures a completion-rate target allows over a stated task volume and window, for example 500 failed tasks a month at a 99.5 percent target on 100,000 monthly tasks.
- Error propagation (multi-agent) Error propagation is the process by which one agent's error becomes trusted input for the agents downstream of it, compounding hop by hop along a chain or fanning out across a topology, so a single local mistake corrupts work far from where it began.
- Exponential backoff with jitter Exponential backoff with jitter is the retry-timing discipline in which each attempt waits longer than the last and the wait is drawn at random from that growing interval, so a population of clients that failed together does not stay synchronized and collide again on a longer cycle.
- Failure attribution (agents) Failure attribution is the diagnostic that traces an observed multi-agent cascade back to the agent and step whose output first went wrong (the root cause), by reconstructing the run as a cascade tree and isolating its earliest corrupted node.
- Failure domain (agent topology) A failure domain is the set of agents, tools and steps that a single fault can reach in a given topology, drawn from the wiring and the shared state rather than measured from a run, so it bounds the propagation radius any one fault can have.
- Graceful degradation (AI agents) Graceful degradation in an agent system is the behavior of returning a reduced but still correct result when a component the run depended on fails, giving up scope, freshness or confidence instead of the whole request, and declaring in the response what was given up.
- Propagation depth Propagation depth is the number of successive hops a fault travels through an agent system before a check contains it, the depth counterpart to the breadth of reach that propagation radius reads.
- Propagation radius (agent systems) Propagation radius, the fault-spread measure sometimes informally called a system's 'blast radius', is the count or fraction of downstream agents and steps a single fault reaches before containment, measured across a defined topology such as an orchestrator with N workers or a chain of depth D.
- Retry storm (agent systems) Retry storm is the failure mode where correlated retries across the layers of an agent stack multiply into far more load than the original traffic, holding a downstream saturated long after the transient fault that triggered the first retry has cleared.
- Timeout budget (agent runs) A timeout budget is a single wall-clock deadline held at the top of an agent run, from which every layer below derives its own timeout by subtracting what has already been spent, so the run is bounded rather than each call being bounded independently.
Reliability testing
3 terms- Chaos engineering for AI agents Chaos engineering for AI agents is the disciplined injection of controlled faults into an agent system to measure how far each fault propagates and what fraction of it the architecture contains, each result reported with a confidence interval.
- Fault injection (agents) Fault injection (agents) is a controlled experiment that introduces a chosen fault at a known point in an agent topology and measures how far it propagates, so its reach and containment become recorded quantities a system can be scored on.
- reliability@k and pass^k pass^k is the probability an agent solves all k runs of one task (closed form p^k). reliability@k is the lane's suite-level aggregate of pass^k: the mean across a representative task suite. It is the consistency counterpart to pass@k (best-of-k capability), not its inverse.
Eval statistics
13 terms- Answer coverage Answer coverage is answers returned over requests sent. Publishing it beside any rate computed on those answers lets a reader see how much of the intended sample the rate actually rests on.
- Bootstrap resampling (eval intervals) Bootstrap resampling estimates the uncertainty of an eval statistic by resampling the scored runs with replacement, recomputing the statistic on each draw, and reading the spread of those values as its sampling distribution. It supplies an interval where no closed-form standard error exists.
- Capability tier (model routing) Capability tier is the band a router sorts a model into, ordered by how much task competence its vendor claims it delivers. The ordering is published as a product hierarchy, so whether a given boundary changes your results is a question only a paired eval on your own tasks can settle.
- Construct validity (benchmarks) Construct validity is the degree to which a benchmark measures the specific capability it claims rather than a proxy a system can score high on without having it; a benchmark is construct-valid only when its top score cannot be earned without the capability it advertises.
- Coverage conditioning Coverage conditioning is the dependence of a published rate on which requests came back with an answer, and it bites when membership of that answered subset correlates with the property the rate is meant to measure.
- Effect size (eval deltas) Effect size is the magnitude of a difference between two eval results, measured on a scale that holds still when the run count changes: on a pass/fail suite, the gap between two pass rates in percentage points, reported with an interval on the delta itself.
- Eval confidence interval An eval confidence interval is the range a procedure produces that, across repeated runs of a suite, brackets a metric's true value a stated fraction of the time (say 95%); its width combines a task-set term (closed-form binomial, or bootstrap) with the seed-to-seed spread, which one run omits.
- Eval reproducibility Eval reproducibility is getting the same result from an evaluation re-run on the same data and the same parameters; it breaks when uncontrolled non-determinism such as sampling temperature, an unpinned seed, or a drifting judge model moves the score while the declared inputs stay fixed.
- Fallback chain (model routing) A fallback chain is the ordered list of models a router tries for a single request, moving to the next entry each time the one before it declines or fails, and ending at the first model that returns an answer or at the end of the list.
- Model router A model router is the component that decides which model handles each incoming request, choosing once per request and before the call is dispatched, so one application can spread its traffic across an expensive tier and a cheap one.
- Refusal rate (LLM models) Refusal rate is the share of requests a model declines to answer on policy grounds, measured over requests sent rather than answers returned. A refusal arrives as a normal response with stop_reason set to refusal, so it never touches an error rate.
- Statistical power (eval design) Statistical power is the probability that an eval reports a significant difference when a regression of a stated size is genuinely present, settled before the run by the drop you would act on, the item count, the score variance, and the false-positive rate.
- Variance decomposition (eval runs) Variance decomposition splits the spread in an eval score into the sources that produced it: which tasks the suite happened to contain, how the model sampled tokens on each attempt, which judge scored the output, and what the harness held fixed between runs.
Judge reliability
8 terms- Cohen's kappa (eval agreement) Cohen's kappa is the chance-corrected agreement between exactly two raters labeling the same items into unordered categories: observed agreement less the agreement their own marginal rates would produce by accident, rescaled by the room chance left. In evals, a judge against a human reference.
- Judge calibration (LLM evals) Judge calibration is the correspondence between the confidence an LLM judge attaches to a verdict and how often verdicts carrying that confidence turn out correct, measured against held-out human labels rather than assumed from the judge's own scores.
- Krippendorff's alpha (eval agreement) Krippendorff's alpha is the proportion by which a set of labels falls below the disagreement chance would have produced, computed for any number of raters, on any measurement scale, with missing labels tolerated. Eval teams use it for judge-against-human agreement on graded rubrics.
- Pairwise vs pointwise scoring Pairwise and pointwise scoring are the two protocols an LLM judge can run under: pointwise grades one response on its own against an absolute scale, pairwise shows the judge two responses to the same prompt and asks which wins. Each licenses a different decision.
- Position bias (LLM judges) Position bias is an LLM judge's preference for whichever answer occupies a given slot in the prompt, first or last, independent of what the answer says. Measure it by scoring each pair in both orders and reporting the share of verdicts that survive the swap, with an interval.
- Rubric drift (LLM judges) Rubric drift is the movement of an LLM judge's effective scoring standard while the rubric text it is sent stays fixed, so two scores produced under the same rubric no longer sit on the same scale. The instrument changed between the measurements.
- Self-preference bias (LLM judges) Self-preference bias is an LLM judge scoring its own model family's responses above the score a reference standard gives the same text, measured as the gap between the judge's own-family win rate and the rate human annotators or a cross-family panel assign to identical pairs.
- Verbosity bias (LLM judges) Verbosity bias is an LLM judge's tendency to score a longer answer above a shorter one of equal or better quality, so response length acts as a scoring feature no rubric named. It is measured by padding an answer with words that add no information and counting the judge's flips.
RAG & retrieval
6 terms- Answer relevancy Answer relevancy is the degree to which a generated answer responds to the question actually asked, scored from the question and the answer alone. It carries no check on whether the answer is true, so it is read beside a grounding score rather than on its own.
- Context precision (RAG evals) Context precision is the share of the chunks a retriever returned that were relevant to the question, scored over the retrieved set rather than over the corpus, and reported as a mean across a query set with an interval.
- Context recall (RAG evals) Context recall is the share of the evidence a correct answer needs that the retriever actually returned, scored per question against a reference for that question rather than against the whole corpus. It sets the ceiling every generation metric downstream is measured under.
- Faithfulness (RAG evals) Faithfulness in RAG evaluation is the proportion of an answer's claims that can be inferred from the retrieved context, computed as verified statements over extracted statements. It scores the generation step against the passages the model was actually shown.
- Groundedness (LLM answers) Groundedness is the share of an answer's factual claims that trace to an identified source a reader could open and check, judged claim by claim against the sources a system named or was handed. It is wider than faithfulness, which scores only against retrieved context.
- Retrieval drift Retrieval drift is the loss of retrieval quality in a live pipeline over calendar time, read as the change between what a frozen query set returns today and what the same queries returned at an earlier baseline. The quantity is a difference, so it carries an interval.