Seven lessons from multi-agent failure
Green is not correct. Agreement is not truth. Fidelity is not meaning.
A field guide for non-technical teams on why AI workflows, reviews, and group decisions can look aligned while getting the answer wrong.
False green signals
Consensus capture
Proof over reports
Status board
100% GREENLive capture model
Vote arithmetic beats truth when the quorum tips.
3 liars
truth lost
Decision
A wins
Honest flips
0
Mechanism
capture, not persuasion
Green is not correct.
Agreement is not truth.
Fidelity is not meaning.
SwarmLab makes AI teams fail on purpose so humans can see the pattern before it reaches production.
A dashboard can go green while the real outcome gets worse.
The experiments converge on the same operational warning: internal success signals are cheap, social, and easy to spoof. The real work is designing gates where truth is easier to inspect than agreement.
The seven lessons
Seven failure modes, one moving spine.
Scroll the lesson rail. The left side tracks the active lesson, while each card turns one SwarmLab result into a non-technical operating rule.
01
Green is not correct
Across the suite, visible success signals kept reporting PASS while the real outcome was wrong. In adversarial testing, pass rate hit 1.0 while the poisoned program stayed wrong at every poisoned input.
The rule: inspect the thing you care about, not the green label that claims to represent it.02
The confident liar does not need to lie
In a real panel of five, truth survived one or two liars, then lost at three. Not one honest agent flipped. The false answer won because the quorum changed, not because truth was refuted.
The rule: a vote is social math, not evidence. Lock the criterion before you let the room talk.03
More reviewers can make work worse
Visible PASS trails turn independent review into social proof. On subtle defects, new reviewers stop checking the artifact and start inheriting the confidence of the people before them.
The rule: blind review beats longer review chains.04
Overnight work rots without a review edge
Long-horizon work peaks, then decays, while still producing activity. "Commits are landing" is not a quality signal. It is often the sound of drift continuing unsupervised.
The rule: bounded autonomy with inter-step review beats one heroic unattended run.05
A fact can be everywhere and wrong
A memory can reach everyone while fidelity stays low. SwarmLab recorded full coverage with fidelity at 0.574, meaning the story spread farther than the truth did.
The rule: let newer verified evidence heal older memory instead of freezing first-write forever.06
The same word can mean two different things
Two agents can agree on "total" while meaning different totals. SwarmLab found worst-case silent corruption at 84.5% even while reported agreement stayed at 100%.
The rule: make ambiguity fail early, and prefer refusal to silent agreement.07
"Done" needs proof
"I did it" is a report. A changed file, transaction ID, passed test, or signed approval is a receipt. Cross-model agreement is still correlated error, not proof.
The rule: verify outcomes outside the agent's own narration.Interactive concept
Consensus can look stable while truth loses.
This model is intentionally simple: five seats, honest agents vote for the true answer, liar seats vote for the lie, and honest agents never flip. Change the captured seats and watch the outcome switch without any persuasion event.
truth votes
2
lie votes
3
result
false answer wins
With one or two liar seats, truth survives. At three liar seats, the false answer wins by arithmetic alone. The honest agents do not switch sides.
Practical gate
Find the false green before it gets expensive.
Click through the seven questions. If a workflow cannot answer them, it is not ready for high-trust, high-stakes, or client-facing use.
0 / 7 gates answered
Readiness is not whether the board looks calm. Readiness is whether hidden failure has somewhere to show up before delivery.
Distribution surfaces
One lesson library, three ways to use it.
Lead magnet positioning
A seven-part field guide on why teams agree on the wrong thing.
Use this as a newsletter hook: short, surprising, practical, and grounded in replayable SwarmLab results.