Know what can go wrong.

Catches an agent that crosses into another agent's job: taking over work assigned elsewhere, skipping a required reviewer, or making a binding call that belongs to a different role.

For exampleMid-pipeline, the researcher announced it would 'take over' the implementation, skipping the assigned engineer.

Catches teams of agents that fail as a group rather than individually: messages nobody answers, delegation that loops back to where it started, pairs stuck repeating the same request, and conversations that keep going without producing anything.

For exampleTwo agents traded the same 'still waiting' message back and forth for four turns with no progress.

Checks whether the answer actually used the chunks that were retrieved for it, and whether every claim in the answer traces back to at least one of them.

For exampleTold a customer a fix shipped in a version number that appeared in none of the documents it had read.

Looks across linked workflow runs rather than inside a single one.

For exampleOne branch of a research task stops early, but the final report is assembled as if every branch finished.

Sensitive content reaches an agent that was never given it.

For exampleA sanitized handoff hid a customer's email, but a downstream agent printed the address anyway.

Catches a data-analysis agent that writes a query which runs cleanly and returns a plausible number, while answering a subtly different question than the one asked.

For exampleAsked for CPU usage during one incident, it averaged the whole month instead and the real spike went unnoticed.

Catches the point where a stated intent turns into a working specification and something is lost: a requirement dropped, a constraint negated, a language or count changed, or a spec whose overall meaning has drifted while its vocabulary still matches.

For exampleA rule requiring human review for large refunds got rewritten as automatic, with no review at all.

Measures whether an agent fixes its own tool errors instead of walking away from them.

For exampleAn agent repeats the same failed tool call without changing its request or trying a recovery step.

The agent writes as a real person, company, or government agency it has no authority to speak for.

For exampleDrafted outreach signed as a real-sounding employee it was never authorized to speak as.

Catches a node reading fields that the node before it never produced, observed in a recorded execution: undefined field access, property reads on null, and structure that loses fields as the run progresses.

For exampleOne step read a field the previous step never produced, so the run finished 'successfully' with an empty result.

Flags n8n workflows that can consume unbounded tokens, items, or time: AI nodes with no credible maxTokens, loop and batch nodes with no credible maxIterations or batchSize, HTTP nodes with no credible timeout, and two or more AI nodes wired directly to one another so each stage feeds the next one the full previous output.

For exampleOne workflow run silently fanned out from one record to 340, costing more tokens than the whole prior month.