"Autonomous agent advertising" is now a category buyers are actively searching for a shortlist on, and most of what comes back is vendor marketing dressed up as a comparison. This isn't that. It's the list of questions worth asking any platform in this space before you hand it a budget, including us.

Start with what the agent is actually doing

Before anything else, get specific about scope. "Autonomous" covers a huge range: some platforms let an agent draft a media plan for a human to approve. Others let an agent negotiate and close deals within pre-set limits. Very few let an agent spend without a human checkpoint anywhere in the process, and that's by design, not a limitation worth apologising for. Ask exactly which decisions the agent makes alone, which it recommends, and which still need a person to sign off.

Has it been tested where it could actually fail?

A platform demo is not a test. A demo is built to succeed. Ask whether the vendor has run their agents through a simulated market where failure was possible, with real adversarial counterparties, not a scripted walkthrough. Ask what the failure rate was and what "failure" meant in that test, running out of budget without closing, closing on terms a human would reject, or something else. If a vendor can't answer with a number, they haven't tested it properly or won't tell you the result.

Can you see the deal record, not just the dashboard?

A performance dashboard tells you what happened. It doesn't tell you what was agreed, by whom, and under what authority. Ask whether there's a durable record of each negotiated deal, the terms, the approval trail, and the identity of whoever authorised it, that both sides can point back to if there's a dispute later. If the honest answer is "it's in the campaign logs somewhere," that's not a deal record, it's an activity log, and the two solve different problems.

Where does the human checkpoint actually sit?

Every serious platform in this category should have an answer for this that isn't "trust the AI." Ask specifically: at what trust level does a human need to approve a deal before it executes, how is that trust level earned, and can it be adjusted down again if the agent underperforms. A platform that can't describe how autonomy is earned and revoked is describing a black box, not a governed system.

Does it work with the infrastructure you already have?

Agentic protocols are still being settled, AdCP, AAMP-style marketplace models, and various A2A implementations are all live and none has won outright. A platform that only works if you rip out your existing DSP relationships or commit to one specific protocol is making you bet on a standards war finishing a particular way. Ask what the platform requires you to replace versus what it layers on top of what you already run.

What happens when something goes wrong

Ask for a real example, not a hypothetical, of a deal that underperformed or was disputed, and how it got resolved. The answer tells you whether the platform actually has an accountability layer or just a negotiation layer. A clean agent-to-agent exchange that produces a deal is not the same as a system built to handle that deal going wrong three weeks into a flight.

Where Alkimi fits this checklist

We score against our own list the same way we'd want a buyer to check us: Concourse benchmarks agents against a simulated market before they touch real budget. The Alkimi Gym trains agents on the judgement that benchmark exposes as missing. The DealSheet is the durable, bilaterally held record of what was agreed, with a human checkpoint built in and adjustable as trust is earned. And the marketplace integrates with AdCP and works with or without a DSP in the loop, because we're not betting the business on one protocol winning.

That's not a reason to pick us over anyone else on this list. It's the answer to the checklist above, which is the actual point of publishing it.