Trellis and Qore vs Amazon's Ads Agent: What a Third-Party AI Does That Amazon's Cannot

Amazon now ships its own AI for ads, so the question for an operator is which jobs to hand it and which to keep. Ads Agent, introduced at unBoxed in November 2025, plans, launches, and optimizes campaigns from inside Amazon's tools, and on September 23, 2026 Amazon gave Seller Assistant always-on workflows and a Selling Partner plugin for Claude and Quick, in beta for US stores.

Both are useful, and both are built around Amazon's own surfaces. If you run one brand on one marketplace and want help inside the console, that may be all you need. If you run a portfolio of accounts, sell on more than one channel, or have margin and inventory rules that decide what a good bid is, the gap between a native agent and a third-party layer starts to matter. This post draws that line, including the places where Amazon's agent is the better pick. It is a Trellis guide, so our bias is on the table: our product Qore is the third-party layer we are comparing against.

Quick answer: Ads Agent wins on first-party access, and a third-party layer wins on scope, control, and logic you own

• Amazon's Ads Agent works with Amazon's first-party insights and needs no setup beyond the console you already use.
• A third-party layer like Qore adds four things a native agent is not built around: reach across accounts and channels, decisions that run on your margin and inventory rules instead of ACoS alone, workflows you can inspect and steer before they run, and actions gated on approval you configure.
• Seller Assistant workflows, as announced, cover Seller Central tasks: inventory, pricing, listings, account health. Amazon has not said they reach advertising, and the plugin is US-only. Treat that as unstated rather than confirmed either way.
• The practical answer for most teams is to use both, and decide job by job which one owns the work.
• The upside and the risk: Amazon's data and your own logic working together, against handing a native agent decisions that depend on costs it cannot see.

Amazon's native AI now covers more of the chat-to-act loop

Any honest comparison starts with what Amazon has shipped, because it is more than it was a year ago.

Ads Agent works inside Amazon's ads tools with first-party insights

Amazon describes Ads Agent as an AI agent that simplifies how advertisers plan, launch, and optimize campaigns. It identifies targeting segments, adjusts pacing across campaigns, and generates SQL for analytics, and it draws on Amazon's first-party insights, including reviewing thousands of audience segments to recommend relevant ones. It is available in Amazon Marketing Cloud and in Multimedia Solutions with Amazon DSP, and Amazon states that campaigns only launch after you review and approve. For background on the clean-room side, see our guide to Amazon Marketing Cloud.

The Ads MCP Server lets outside AI tools reach Amazon Ads

On February 2, 2026, Amazon opened its Ads MCP Server to global open beta for Amazon Ads partners with active API credentials. It turns natural language prompts into structured API calls, so an agent can create, update, or delete campaigns, run reporting queries, manage account-level settings, and access billing data. Amazon lists Claude, ChatGPT, and Gemini as connectable platforms. We cover what that protocol means in practice in What Is MCP for Amazon Sellers and Can Claude Manage Amazon Ads.

MCP, the Model Context Protocol, is a shared way for an AI assistant to call an outside service's tools. The Ads MCP Server is Amazon publishing that door for Amazon Ads: an assistant like Claude can reach the same campaign, reporting, and billing actions an advertiser reaches by hand, without a custom integration built for each one. That is the single most consequential line in this whole comparison for a third-party layer, because until this door existed an outside agent could only work from exports a person pulled. Now it can read and act in the account directly.

Two limits keep it from being the whole story, and both point at where the third-party value actually sits. The server is a pipe, not a brain: it exposes the actions, and it does not decide which action is right for your margin, your inventory, or a second marketplace it cannot see. And it reaches Amazon Ads only, so a third-party agent still brings the cross-channel scope, the per-account standard, and the approval gate itself, and it reaches every other channel through that channel's own API, not this one. It also does not hand over the privileged first-party signals Amazon's own Ads Agent draws on, which is the asymmetry the rest of this piece turns on. So the door Amazon opened is what lets a layer like Qore reach Amazon at all, and everything that makes the layer worth running sits on top of the door, not in it.

Seller Assistant workflows cover Seller Central, and Amazon has not said they cover ads

At Accelerate on September 23, 2026, Amazon added always-on workflows to Seller Assistant and a Selling Partner plugin that brings seller data into Amazon Quick and, in beta for US stores, Claude. The examples and templates Amazon published cover ratings, account health, orders, pricing, restocking, listings, and product expansion, and Amazon describes the plugin as connecting listing contributions, inventory information, performance metrics, and sales analytics. We covered the launch and its open questions in our Seller Assistant workflows post.

Two things stand out for an advertiser. First, advertising is absent from what Amazon has said. Nothing in the announcement or the trade coverage describes a Seller Assistant workflow reading a campaign or moving a bid, so as of today the safest reading is that these workflows are not connected to ads. That could change, and Amazon may simply be routing ads work through Ads Agent instead, but it is a notable gap: advertising is where a large share of Amazon's own revenue comes from and where sellers spend the most hours, and the flagship seller automation launch of the year does not mention it. Second, the plugin is US-only for now, so a seller outside Amazon's US stores gets the workflows inside Seller Central and not the Claude or Quick route.

The workflows look strongest for automating Seller Central tasks Amazon already understands: a rating alert, a restock plan, a price response inside one account. Where the announcement goes quiet is on the mechanics an ads or cross-channel workflow would depend on. Amazon has not specified:

• the maximum complexity of a workflow, or limits on how often one can run
• trigger types beyond a schedule or an on-demand run, such as a rule that fires when ACoS stays above a threshold for two consecutive hours
• how much history a workflow can read, and whether that window can be changed
• whether workflows can be extended through an API or only through the interface
• whether a workflow can coordinate across marketplaces or apply across more than one selling account
• what happens when a step fails: retries, halts, or a silent skip
• role-based permissions on who can create, edit, or approve a workflow

None of that means the answers are bad. It means that a seller building an operating standard on top of these workflows is building on terms that have not been published, and an ads workflow in particular has no stated home there yet. The differences that remain between Amazon's native tools and a third-party layer are about scope, control, whose logic runs, and what the workflow can see.

Four things a third-party layer like Qore does that a native agent is not built to do

Qore is the workflow layer of Trellis. You describe a data or action workflow to Q, our assistant, and Q assembles it as a visible graph of steps. You review it, lock it as a skill, and it runs on a schedule or on demand. Here is where that model differs from a native agent.

It works across accounts and across channels

A native agent is built around one marketplace's surfaces. Qore connects Amazon, Walmart, Google, Shopify, and TikTok, and a locked skill can be templated across a roster of accounts. For an agency or a brand with regional accounts, that means one negative-keyword review or one budget-pacing check written once and run everywhere, with the output landing in Slack, email, or a spreadsheet where your team already works. A Walmart budget question and an Amazon budget question can sit in the same workflow instead of two separate chat windows. The inputs are wider than marketplace data too: a Qore skill can read the connections you give it through MCP and deliver to Slack, email, or a spreadsheet, so a rule can weigh a supplier lead time or a cost sheet that no Amazon report carries. Seller Assistant workflows, as described, run on what Amazon holds; the Claude plugin puts that data next to a seller's other connections, but the workflow itself does not read them.

It decides on your margin and inventory rules, not ACoS alone

An ACoS target treats a 40 percent margin SKU and a 12 percent margin SKU the same way. Operators do not. In Qore you write the rule into the skill: pause spend on ASINs below a set days of cover, hold bids where margin after fees does not support the ACoS, treat a launch SKU differently from a mature one. Qinetix, our ads and dynamic pricing platform, already reads live Buy Box status and inventory health as part of how it optimizes. Seller Assistant now reaches inventory too, so the useful test is specific: ask any tool what data its ad recommendation read, and whether your own cost and margin rules were part of it.

It shows its logic and lets you steer it before it runs

A chat answer from a language model can differ from one run to the next. A Qore skill is built with the model, then locked. After the lock there is no language model at runtime, so the same inputs produce the same outputs, and you can open the graph and see which data source, which filter, and which threshold produced a result. If a rule is wrong, you edit the step. That is the difference between an agent you supervise after the fact and a workflow whose logic you own. It is also why we say Q helps you build a skill and Qore runs it.

It gates actions on approval, and you set the gate

Amazon's Ads Agent requires review before a campaign launches, and that is the right default. Qore applies the same principle at a finer grain. Actions such as bid changes, pauses, negations, and budget moves default to manual approval, you configure which actions need sign-off, and you can let a trusted action run automatically once it has earned that. Approval on a locked skill means you are approving the output of logic you have already inspected, not a fresh generation each time.

Side by side: a native agent and a third-party layer

Where Amazon's Ads Agent is the better choice

Use it when the job depends on data only Amazon holds. Audience segment recommendations built on Amazon's first-party insights and SQL generation inside Amazon Marketing Cloud are the clearest examples. A third-party tool cannot see past what the APIs expose, and pretending otherwise would be a bad pitch. It also wins on speed to first value: there is nothing to build, and a small team with one account can get real help on day one.

How to split the work: a ladder from native to codified

Most teams do not need to pick a side. A sensible progression looks like this:

• Start with Ads Agent for planning, targeting suggestions, and AMC analysis inside Amazon's tools. You keep first-party access and give up nothing.
• Add a general-purpose assistant through the MCP Server if you want to query your account in plain language. You gain flexibility and give up repeatability, because each answer is generated fresh.
• Codify the routines you repeat in a skill once you notice you are asking the same question every Monday. That is where Qore fits: the review you run by hand becomes a workflow that runs on schedule, across every account, on rules you wrote.

Our own view of where general-purpose assistants fall short on ads work is in AI for Amazon Ads.

The honest limits of Qore

Qore is in open beta, and features are still being added. Its insight and analysis stand alone, but the action loop assumes Trellis is your bid manager of record, so it is a fit for teams willing to run bids through Qinetix. Q helps you build and shape a skill, but it will not decide your strategy: the margin floor and the inventory threshold have to come from you. And a locked skill is only as good as the standard behind it, which takes some time to write down properly.

Conclusion: let Amazon's agent own Amazon's data and own your logic yourself

Amazon's native tools are getting better fast, and that is good for every advertiser. The work that stays yours is the part no marketplace can hold for you: how you weigh margin against growth, which channels share a budget, and what has to be true before money moves. If that logic lives in a chat window, it lives nowhere. If it lives in a skill you can open, edit, and approve, it compounds. To see how a locked skill looks in practice, take a look at Qore or book a walkthrough.