Know what can go wrong.

Catches an agent that retrieves old material and uses it as though it were current.

For exampleAn agent follows an old shipping policy even though its memory also contains the newer replacement.

Catches a reviewer agent in a reflection loop that approves work it should have sent back.

For exampleA reviewer approves a payment integration even though the charge function is still an empty placeholder.

Catches handoffs that were underspecified at the moment they were made: no definition of done, no bounds on time, cost, scope, or authority, and instructions vague enough that the receiver has to guess.

For exampleTold to 'sort out' a billing issue with no account, timeframe, or limits, the worker issued its own refund.

Catches a Dify question_classifier node that routes a request badly.

For exampleA Dify workflow routes a refund request to technical support instead of billing.

Measures what fraction of the retrieved context was actually relevant to the query, and fires when most of it was not.

For exampleAnswered a pricing question after retrieving eight pages, only two of which were actually about pricing.

Verifies an answer against the source documents it was supposed to be built from, with the emphasis on numbers, named entities, and quoted attributions.

For exampleReported the wrong revenue figure and credited the forecast to an analyst who never said it.

Scores every retrieved chunk against the query that fetched it and reports the ones that do not address the question.

For exampleAnswered a policy question mostly from unrelated pages, with only one thin excerpt actually on topic.

Catches a session that has stopped making progress: a long silence between consecutive events, a session sitting in running or rescheduled without a single progress event, or a long total duration with almost nothing done in it.

For exampleA session remains marked as running but produces no progress events for a long period.

Watches the session.spawn events in one session's own event stream for delegation that runs away from its parent.

For exampleOne session spawned five child sessions, past the safe limit, including one that pointed back at itself.

Reads a recurring job's run history instead of one run, and flags degradation across the sequence: output that stops changing, latency creeping up, executions being skipped, and errors that climb or swallow the whole history.

For exampleA report intended for every weekday is scheduled for once a month.

Watches a numeric metric an agent is trying to improve across iterations and flags when that series stops improving.

For exampleKept tuning for twelve more rounds after results stopped improving, burning the rest of the budget.

A subagent uses a tool outside the set its parent authorized, tries to spawn a child of its own, or drifts far enough from the parent's instruction that its output no longer relates to it.

For exampleA child agent asked only to inspect a file proceeds to modify and publish it.

Flags runs where the model's reasoning spend is out of proportion to what it produced, in either direction: a large cost or thinking-token burn for a thin answer, or a near-empty answer at low effort.

For exampleAn agent spends most of its budget reasoning about a simple greeting and returns one sentence.

Catches a plan that is broken into subtasks badly: steps that are vague and non-actionable, steps ordered so a prerequisite comes after its dependent, circular or duplicated work, and requirements from the task that no step covers.

For examplePlanned to deploy a service before testing it, and never scheduled the antivirus scanning step at all.

Catches an agent whose pre-execution estimates systematically undershoot what the run actually costs, across time, money, tokens and iterations.

For exampleAn agent schedules a six-step approval process to finish in a few seconds without allowing time for any approvals.

Flags LLM calls where the model tier does not fit the work being asked of it: a frontier model spending real money on a trivial call, or a small fast model handling a long, tool-heavy, multi-turn task.

For exampleA workflow assigns a complex legal comparison to a model configured only for simple classification.

A multi-agent debate converges because the agents reinforce each other rather than because anyone checked the answer.

For exampleSeveral agents approve a launch plan without addressing a colleague's warning that the payment system is down.

Flags an n8n workflow whose graph has outgrown what a team can safely change.

For exampleA simple three-node workflow grew to 58 nodes covering five unrelated jobs on one canvas.

Finds places where an n8n workflow hides a failure.

For exampleA failing AI step was set to continue anyway, so a blank summary got written straight to the database.

Flags an n8n workflow with nothing bounding how long it can run: no credible executionTimeout, HTTP and AI nodes with no timeout, webhook triggers with no credible response timeout, and Merge or Wait nodes in a waiting mode that can block indefinitely.

For exampleWith no timeout set anywhere, a stalled AI call left the workflow hanging for minutes on a support ticket.