Every agent demo you have ever seen was a success case. The agent reads the board, writes the summary, everyone nods. That's not dishonest — it's just not the part you need to prepare for.
Here's the part you need to prepare for. An agent writes your sprint summary. It's well structured. The tone is right. Every sentence in it is true. And it doesn't mention that the integration work is blocked, because the blocker was recorded as a comment on a subtask rather than as a status change, and the agent didn't look there.
You skim it. It reads fine. You send it to your stakeholders.
Nothing about that output looks wrong, because nothing in it is wrong. That's the failure mode. Not error — omission, delivered fluently.
Omission. The most dangerous and the least discussed. Wrong sentences get caught; missing ones don't. Nobody proofreads for absence. And the things most likely to go missing are exactly the things recorded off the happy path — a blocker in a comment, a dependency in a linked ticket, a concern someone raised verbally in standup.
Fabrication. Ask for something the data doesn't contain and you may get it anyway, complete with plausible names and dates. Practitioners in the Atlassian community have reported agents producing entirely fictional releases when the underlying data had nothing to offer. The invented material is not flagged as invented. It looks exactly like the real material.
Stale confidence. The agent answers from whatever slice of the board it pulled, with no sense that the world moved. In a live sprint, "as of some point earlier today" and "now" are different answers, and only one of them is useful in a stakeholder meeting.
The instinct is to add "tell me what you're unsure about" to the prompt. It feels responsible. It doesn't work — the model's self-reported uncertainty doesn't reliably track where it's actually wrong. You get a confident answer with a confident-sounding caveat attached, and you've learned nothing.
You cannot ask the thing that was confidently wrong to notice it was confidently wrong. Verification has to come from outside the output.
These don't make the agent right. They make checking cheap, which is the thing that determines whether you actually check.
1. Scope hard. Current sprint only. This board. This label. A narrow question has a narrower space to be wrong in.
2. Make it cite. Every claim tagged with the ticket ID it came from. This is the highest-leverage one — it turns verification from "reread everything and hope" into "click three links." Anything that can't be sourced becomes visible immediately.
3. Make it declare absence. "List anything you looked for and could not find." This is the direct counter to the omission failure, and almost nobody does it.
4. Make it flag inference. "Mark anything you inferred rather than read." Separates what's in Jira from what merely sounded plausible.
"Review all AI output" is a policy that lasts about three sprints. Then a release goes sideways, everyone's busy, and it quietly stops. Reviewing everything is how you end up reviewing nothing.
Sort by what a wrong answer costs instead:
A retro summary that's slightly off wastes five minutes. A capacity forecast that's confidently wrong plans your next sprint into a wall.
Three properties separate the review habits that last from the ones that die in March.
Structural, not diligent. If it depends on remembering to be careful, it decays. If it's a step in something that already happens, it survives.
Attached to an existing ceremony. The sprint report gets checked in the five minutes before review, because that slot already exists and has a forcing function. A standalone "AI review checkpoint" on your calendar will be gone by the third sprint.
Author separated from verifier. Whoever ran the prompt is the worst person to catch its errors — they already believe it. Where the stakes justify it, someone else looks.
Short list, and it's short on purpose.
Anything you'd have to defend under scrutiny without being able to say where the number came from. Anything about a specific person — performance, capability, blame. Anything where the failure is discovered by someone other than you, months later.