Agent Negotiation
Put two agents with different instructions on the same problem and they will disagree — one is told to protect a budget, the other to secure the earliest delivery date. Agent negotiation is the structured exchange through which they reach terms neither could have set alone.
How a negotiation is set up
The design work is in the framing, not the dialogue:
- Each agent gets private information and an objective — a reservation price, a deadline, a quality floor.
- A protocol fixes whose turn it is, what counts as a valid offer, and what an acceptance commits to.
- A termination rule — agreement, a round limit, or a walk-away threshold — keeps the exchange finite.
Implementations run this as role-conditioned agents exchanging messages in AutoGen or CrewAI, sometimes with a third agent adjudicating.
What it is good and bad at
Negotiation surfaces trade-offs a single prompt would quietly resolve on its own: a procurement agent arguing with a supplier agent exposes which constraint actually binds. It fails in recognisable ways too. Agents sharing one model drift towards agreement too readily, producing consensus that reflects politeness rather than evaluation. Rounds multiply token cost with no guarantee of movement. And an agent that treats its counterpart's claims as facts can be walked anywhere.
Why it matters
When a task spans parties with genuinely opposed interests — buying, scheduling, allocating scarce capacity — negotiation is the coordination pattern that fits. It needs the same spend caps and audit trails as any other loop, with one twist: here the step count is set by how stubborn the other side is.
Agent negotiation is an interaction pattern in which agents with differing goals or constraints exchange proposals until they reach an agreement — a price, an allocation of tasks, a schedule. Each agent argues from its own objective rather than a shared one. Negotiation lets independently owned agents coordinate without a central authority, and the protocol governing offers and acceptance determines whether the exchange terminates at all.
Think of buying a second-hand car. You want a low price, the seller wants a high one, and you go back and forth until you meet somewhere in the middle. Agent negotiation is two pieces of AI software doing that on behalf of different sides.
Each one is given a goal and some private limits — the most it may spend, the earliest date it can accept. Then they exchange offers in turn until they agree or walk away.
The appeal is obvious for booking, buying and scheduling, where the haggling is tedious and nobody enjoys it. The catch is that software can be talked round too, so people set firm boundaries on what it is allowed to agree to.