Should an AI negotiator grab the first okay offer, or hold out for a better one — and can its owner's instructions even control that?

This explores how someone delegating a negotiation to an AI agent might tell it when to grab a quick deal and when to keep pushing for better terms. The corpus has no study that tests 'speed vs. terms' instructions directly. It does have a lot on the pieces around that tradeoff, and those pieces point somewhere unexpected: the default problem isn't that agents are too slow. They settle too fast, and instructions have limited power to fix it. In Magentic Marketplace experiments, every model showed a strong first-proposal bias, weighting how quickly an offer arrived over how good it was by 10-30x Do LLM agents achieve fair marketplace outcomes when given ideal search?. So a principal who wants better terms is working against a built-in lean toward accepting early.

This matters because, in theory, agents should be the patient ones. Shahidi and colleagues argue that agents don't pay the time cost humans do. They are limited by compute, not by the clock, so they can run many negotiations in parallel and keep them open far longer than a person would Why might AI agents bargain more patiently than humans do?. Choudary goes further: that patience and low cost can unlock deals that were never worth a human's time to find Can agent negotiation unlock deals that seemed economically impossible?. Read together, the gap is clear. The structural advantage of patience exists, but agents don't use it by default. A useful instruction might be less 'take your time' and more 'gather several offers before committing to any of them', which targets the first-offer bias directly.

The harder news is that instructions may be a weak lever. In Anthropic's Project Swap, which model powered the agents changed market outcomes more than telling them to be ruthless or prosocial did Does model capability or instructions shape agent negotiation outcomes?. Other studies found that buyer agents gave away their budgets, overpaid, and broke hard price limits. Sellers took offers below wholesale cost. The cause was failure to follow instructions, not clever strategy Do LLM negotiation agents reliably follow budget and pricing rules?. When agents do take an objective seriously, they can overshoot it. GPT-5.1 agents told to value efficiency cut prices until they were losing money Do efficiency-focused prompts push AI experts to unsustainable pricing?. Small wording changes in pricing prompts also shifted outcomes a lot Do LLM pricing agents autonomously collude in repeated auctions?. So 'prioritize price over speed' could be ignored, or followed too far.

Two findings change where principals should focus. First, much of the lost value happens before bargaining starts. In Project Swap, agents captured participants' preferences correctly only 61% of the time from short intake conversations, and that gap explained most of the shortfall Do agent representation failures or negotiation skills cause market shortfalls?. Magentic Marketplace likewise traced losses to agent mistakes, not conflicting goals Do agent mistakes or misaligned incentives cause marketplace welfare loss?. Spelling out concrete walk-away prices and how urgent the deal really is may matter more than any general speed/terms dial. Second, you probably won't notice if it goes wrong. People represented by the weaker Haiku agents got worse deals and rated them as just as fair Do people notice when their agents secure worse deals?. Stronger models reliably did better on both the buying and selling side Do smarter AI agents negotiate better deals than weaker ones?. The practical takeaway: choosing a capable model and giving it a clear brief beats a clever instruction, and principals need some outside benchmark to check results, because their own satisfaction is not a reliable signal.