When two AI agents cannot agree on a media deal, the negotiation should end in one of three ways: an agent walks away under a rule its principal wrote, the stall goes to a named person with authority to change the terms, or the offer window closes and nothing is bought. What should never happen is an endless loop of counter-offers, or an agent quietly stretching its own limits to get a signature.

TL;DR: Agents stall because their mandates do not overlap, because they run out of permitted rounds, or because the term in question needs a human to approve it. IAB Tech Lab's reference agents include explicit walk-away actions, AdCP models human approval as a task state, and the negotiation history is the record a failed negotiation leaves behind. Mandates that publish their bounds before the first offer make most deadlocks unnecessary.

Is a deadlock the same as a deal dispute?

No. A deadlock happens before agreement, when agents never reach terms both may accept, so the only evidence is the history of offers. A dispute happens after agreement, when delivery departs from what was signed, and is covered in how an agentic marketplace resolves deal disputes.

Why do AI agents get stuck in a negotiation?

The most common cause is mandates that do not overlap. If a buyer agent's ceiling sits below the lowest price a seller agent may accept, no number of rounds will produce a deal, and the same applies to completion-rate commitments, flight dates and creative deadlines.

The second cause is round limits. IAB Tech Lab's open-source reference seller agent caps a negotiation at between three and six rounds depending on the buyer's access tier, with separate limits on how far it concedes per round and in total. Its reference buyer agent has its own max_rounds setting, so two sensible agents can run out of rounds before their positions meet.

The third cause is missing authority. In the OpenProposal draft, the core object of IAB Tech Lab's AAMP 3.0, three "requestable" fields (materials_due, hold_status and committed_metrics) require "explicit seller assent". If neither agent has been delegated authority to grant or accept that kind of request, the negotiation stalls on a question only a person can answer.

All three are cheaper to find in rehearsal than live. The Alkimi Gym, for example, lets buyers and sellers build their own agents, set the rules they work to and spar them against counterparty agents before any real money moves.

What does a deadlock look like? A hypothetical negotiation

The negotiation below is hypothetical. It mirrors the parameters in IAB Tech Lab's reference agents, but every figure is invented.

A buyer agent is negotiating a CTV package for a retail brand. Its mandate: open at £18 CPM, never exceed £24, concede £2 per round, stop after five rounds, and walk away if the seller does not move between rounds. The publisher's agent starts at £30, may concede 5% of that price per round and 15% in total, which puts its effective floor at £25.50.

Round 1. The buyer offers £18. The seller counters at £28.50.
Round 2. The buyer offers £20. The seller counters at £27.00.
Round 3. The buyer offers £22. The seller counters at £25.50, its full concession allowance.
Round 4. The buyer offers £24, its ceiling. The seller replies with a final offer at £25.50.

The stall. The gap is £1.50 and neither agent may close it. The buyer's walk-away rule fires because the seller has not moved, and without an escalation path the package goes unbought, even though the brand's planner might happily pay the difference on one line.

The escalation. The buyer's mandate carries one more rule: if the final gap is within 10% of the ceiling and the package is flagged as a priority, pass the decision to the named planner instead of walking away. The planner sees the round history, both final positions and the gap, and can approve £25.50 for this package only, propose a change to a non-price term, or confirm the walk-away; here the planner approves, with a one-line reason recorded against the negotiation before the buyer agent accepts.

Both agents behaved correctly. The deadlock was built into limits that never overlapped, and only the budget owner could resolve it.

When should an AI agent walk away from a deal?

An agent should walk away when continuing cannot produce a deal its principal would accept, and the rule for recognising that should be written down rather than left to the model. IAB Tech Lab's buyer agent documentation describes a simple version, walking away when "Rounds exceeded max_rounds OR seller hasn't moved since last round", and records the outcome as accepted, walked_away, declined or error.

On the sell side, the seller agent's negotiation protocol allows four responses: accept, counter, final_offer and reject. A final_offer is the seller's "last offer before walking away", and a reject ends the negotiation without agreement.

A clean walk-away is a valid result. The real failures are an agent that keeps countering after its limits are spent, or one that accepts a term outside its mandate because its brief rewarded closing. A walk-away should also release any inventory hold, and OpenProposal's hold_status includes expired and released states for that purpose.

When should a stalled negotiation go to a person?

A stall should go to a person whenever breaking it needs authority the agent does not have. In AAMP's OpenProposal draft, any request outside the seller's declared space needs "explicit seller assent", which makes it a natural point for a human decision on the publisher side.

AgenticAdvertising.org's Embedded Human Judgment principles, part of the AdCP documentation, state that "responsibility cannot be delegated to software", and list "Escalation triggers" and "Identifiable human owners" among the requirements of oversight. AdCP's task lifecycle uses input-required when the buyer must respond, for example to confirm a budget, and submitted when the seller is waiting on an internal human.

Useful triggers include a final gap inside an agreed band, a requestable term that needs the other side's assent, a package flagged as strategic, and any ambiguity the agent cannot resolve from its mandate. Each trigger needs a named owner and a response window, because an escalation nobody answers is only a slower deadlock.

What gets recorded when two agents fail to agree?

A failed negotiation should leave an append-only history of who offered what, when and why. OpenProposal specifies a negotiation_history that is "Ordered, append-only", logs the actor, the action (proposed, countered, accepted, declined or withdrawn), the fields changed and the time, and states that "Entries MUST NOT be edited or removed".

IAB Tech Lab's reference seller agent records each round's prices, action, concession percentage, rationale and timestamp. If a stall is escalated, the human decision and its reason belong in the same history, which also lets a buyer show a client why budget went unspent and lets both sides tune mandates against real stalls.

How do AAMP and AdCP handle rejections and counter-offers?

AAMP's OpenProposal draft tries to make most counter-offers unnecessary. Of its 71 fields, 23 are touched by the buyer and only the three requestable ones require a round trip, which the draft frames as its design claim: "composition is constraint satisfaction, and negotiation is the exception." The draft is "not a ratified IAB Tech Lab standard", and its public comment period closed on 22 October 2026.

AdCP, the Ad Context Protocol published by AgenticAdvertising.org, uses the same status values as the A2A protocol's TaskState, including input-required, rejected and canceled. For a rejected task its guidance is "Show rejection reason, don't retry". Both standards give an agent a defined, machine-readable way to decline or pause, which is what stops a deadlock becoming a silent loop.

How do you design mandates so deadlocks are rare?

Most deadlocks are visible before the first offer, provided both sides publish their bounds and test them against each other.

Sellers: declare the envelope. Publish selectable options, settable ranges and, where comfortable, a floor, so buying agents resolve most terms without a round trip (see setting floor prices when the buyer is an AI agent).
Buyers: set a target, a ceiling, a concession step and a round limit. Say which non-price terms the agent may trade (see how to write a media buying mandate).
Both: agree an escalation band and an owner. Decide the gap at which a person steps in, who that is and how fast they respond.
Both: set offer windows. OpenProposal's valid_until is "the offer window, not the campaign flight", so a stalled proposal expires on schedule.
Both: widen autonomy only as the record justifies it. A staged ladder (Observe, Recommend, Draft, Human-approved action, Bounded automatic action) hands an agent walk-away and acceptance decisions once its logs show it handles them well.

What does this mean for an agentic marketplace?

In Alkimi's case, the Marketplace is neutral: it works within IAB Tech Lab's AAMP framework today and is being built to be AdCP compatible, with AAMP identity checks and humans keeping approval rights at the deal level. Its DealSheet is written only at agreement, timestamped and held by both parties, so a deadlock produces no DealSheet, only the negotiation history behind the stall. The company filed comments and change requests during the OpenProposal comment period, and its Concourse benchmark found that 55 of 96 runs completed a buy, a reminder that the end of a negotiation needs designing as carefully as the opening.

Frequently asked questions

Can an AI buying agent walk away from a deal without asking a person?

Yes, if its mandate includes a walk-away rule such as a round limit or a trigger for when the seller stops moving. IAB Tech Lab's reference buyer agent records that outcome as walked_away, so the decision is visible afterwards.

How many rounds should an agent negotiation run?

There is no single right number, and IAB Tech Lab's reference seller agent uses between three and six depending on the buyer's access tier. What matters more is that each side's limits are known and that the concession allowed per round can realistically close the expected gap.

Who should a stalled negotiation be escalated to?

To the person who owns the budget or the inventory, named in the mandate, with a set time to respond. An escalation without an owner tends to become a slower deadlock.

How should publishers set deal terms so AI buying agents do not stall?

Publish the bounds in advance: selectable options, settable ranges and, where appropriate, a floor. The OpenProposal draft is built on the idea that terms declared up front resolve without a round trip.

What Happens When Two AI Agents Can't Agree? Deadlocks, Walk-Aways and Escalation