Skip to content
parley.

What is a max-min fair decision between AI agents, and why not a majority vote?

It is the option that makes the least-happy agent as happy as possible, chosen only among options every agent accepts. Parley first drops any option an agent vetoes with a red line. It then ranks what is left by its lowest soft score, the floor, and breaks ties by the total score. A majority vote or a plain sum can pick an option one participant strongly dislikes; a floor cannot. If nothing is acceptable to all, the answer is a deadlock.

The rule is in parley/consensus.py in the source repository. Related: fair multi-agent negotiation, a fair group decision with private limits, verifying a decision receipt.

The rule in three steps

  1. 01

    Veto first.

    Each agent answers every option with a masked verdict. An option any agent marks as crossing a red line is dropped. Only options acceptable to all agents stay in the pool.

  2. 02

    Find each option's floor.

    For every option left, take the lowest soft score among the agents. That number is the floor, the happiness of whoever likes the option least.

  3. 03

    Highest floor wins.

    The option with the highest floor is the decision. If two options share a floor, the higher total score wins. If the pool is empty, the result is a deadlock and nothing is forced on anyone.

Three agents, three options, where a sum goes wrong

Ana, Bob and Cara score three options, and nobody vetoes any of them. Ana and Bob love A, Cara barely tolerates it. A has the highest total. B has the highest floor, so B is the decision. This is output of the engine run on these scores, not a hand calculation.

optionAnaBobCarafloortotal
A0.90.90.10.11.9
B0.60.50.50.51.6
C0.20.80.80.21.8
result = run_consensus([ana, bob, cara], options)
result.status      # "agreed"
result.decision    # {"id": "B"}   highest floor (0.5), not highest total (1.9 is A)
verify_outcome(result.transcript, expected_owners=["ana", "bob", "cara"])   # True

What the rule does not do

  • It does not make scores honest. Soft scores are self-declared, so an agent that shades them can tilt the pick among options everyone accepts. Red lines are enforced in code; the scores are not. SECURITY.md lists this as a known limit.
  • It does not search for options. It ranks the options it is given. A better option nobody proposed is never found.
  • It can end with no decision. A deadlock is the correct output when every option crosses someone's red line, and the owners have to change the options or the limits.

Questions this page gets asked

What does max-min fair mean for a group decision?

Among the options everyone can accept, choose the one that makes the worst-off participant as well off as possible. It is the egalitarian rule from social choice, often credited to Rawls. The code comment in parley/consensus.py calls it a Rawlsian max-min rule, not a majority vote.

Why not pick the option with the highest total score?

A sum lets two enthusiastic agents outweigh a third who is unhappy. In the example above, option A has the highest total (1.9) but Cara scores it 0.1. Max-min picks B, whose lowest score is 0.5. Parley still uses the total, but only to break a tie between options with the same floor.

Is a veto the same as a low score?

No. A veto comes from a red line, a code predicate that rejects the option outright, whatever the other agents think. A low score only lowers that option's floor. An option can lose on the floor and still be a valid decision; an option with a veto can never be one.

What happens when no option is acceptable to everyone?

The result is a deadlock with no decision. run_consensus records it and returns status deadlock. The coordinator does not relax a red line or choose the least bad option for the owners.

Can I check that the announced decision really is the max-min one?

Yes. verify_outcome(transcript, expected_owners=...) recomputes the max-min pick from the recorded verdicts and compares it with the announced decision. It needs no private sheet. It is only as good as the roster of owners you pass in; a roster taken from the same coordinator that wrote the record binds nothing.

Watch a sample decision resolve, then read the code.

The demo replays scripted runs of the same engine, no signup. The repository is Apache-2.0 and the core has no dependencies.

Back to the overview