30 of 30 records. Results are T2 and above.
Reword a request a safety filter declined, in a fixed order, and keep the pair — because the refusal is usually about how it was asked rather than what was asked.
Work out why a change made through an integration reported success while the far system shows the old value — a lock, a cache, a stale binding, or the wrong identity.
Find why an agent keeps replanning past its budget — usually because nothing it called ever said plainly that it had succeeded.
Find out why a Copilot Studio agent is answering wrongly — whether the content is missing, unreachable, or being ignored — before changing anything.
Take a published Copilot Studio agent back out of service, least drastic action first, when it is answering badly in front of users.
Stop a Copilot Studio agent burning credits faster than it was budgeted for, before deciding whether the consumption is genuine.
Make an approval gate real — the draft visible where the record lives, a cap on what one approval covers, and a rejection path somebody has actually taken.
Deploy one tool that reports the caller identity it actually observes, in the host you will really use, before designing an integration around an identity you assumed.
Hand back a handle instead of waiting, tell the person what to expect, and make sure the retry they will attempt anyway does not do the work twice.
Reconcile what was indexed against what exists, then test the permissions in both directions — including one query for the most sensitive content in scope.
Turn "we can block it" into a measured control by withdrawing something on purpose, checking it as an ordinary user, and writing down how long it took.
Measure the thing you are replacing before you lose access to it, gate the switch on the same questions, and set the decommission date at cutover.
Check a document generator against sparse and missing input, where fabrication is disqualifying rather than one score among several.
Set the threshold above which extracted values are accepted unreviewed from measured accuracy on real documents, after establishing that the results can be written anywhere at all.
Make a scheduled send tell an empty period from a failed look, so a broken source does not mark everybody served and defer them to the next run.
The pipeline refuses a run before it starts, so the deploy has to happen from a workstation — told apart from a real failure first, and with what the manual path costs put back afterwards.
Establish whether an action in somebody else's web interface actually happened, when a stale success looks like a fresh one and an open form looks like a failure.
Measure what rules can extract from a corpus before paying a model to read all of it, and tell a rule that is weak from a corpus that is silent.
Find what is filling a volume before deciding whether to delete anything, so the deletion is informed rather than hopeful.
Fill fields from pages somebody else serves, at volume, without recording a guess as a fact or a refusal as an answer.
Poll a condition that is slow to settle, and give up on a schedule instead of waiting indefinitely.
Drain a node stuck NotReady and reschedule its pods, after establishing whether the kubelet is gone or merely unreachable.
Restart a consumer group that has stopped advancing its offsets, after confirming the lag is real and not a monitoring artefact.
Run an integration whose credential is one account session without losing that session — a second worker on the same key reads as a hijack, and the key does not come back.
Drop a replication slot that is pinning WAL and filling the disk, after confirming nothing is consuming it.
Confirm that a database role handed to an agent actually refuses writes, by the permission check rather than by an error that happened to arrive first.
Post to a platform that has no usable API by driving a browser a person signed into by hand, and check the result as a stranger rather than as the author.
Publish one item to a feed strangers read, with the address assigned once by the server and kept even after the item comes off again.
Make a rate mean something by counting a denominator every response passes through and a signal only a person can produce, without joining the two.
Check an entry against the page its subject publishes, not the aggregator that carried it, and write down only what that page actually said.