It isn\u2019t a \u201cmagic\u201d bridge that gives AI free rein. It\u2019s a curated set of capabilities defined by the underlying tool. The agent only sees the data and takes the actions it has been explicitly given permission to handle.<\/p>\n
![\"\"](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/09\/01-claude-cowork-custom-connector-mcp-420.jpg\")
Screenshot by author, August 2026<\/p>\n
With connectors, your agent can go beyond talking strategy with you; it can also implement the changes you agree on.<\/p>\n
This makes it possible to start working in a whole new way and it\u2019s radically changing what people expect from software. They\u2019ve stopped wanting to learn interfaces. They want to state an outcome and have something competent go do it.<\/p>\n
3 Ways To Work, And 2 Versions Of Each<\/p>\n
Today, I see three core ways people use software like Google Ads, and the two new ways foreshadow the end of the day we log into the Google Ads UI to do work in PPC.<\/p>\n
1. Working By Hand In The UI<\/p>\n
In the Google Ads UI, you get ground truth, and you see what exists rather than only what you thought to ask about. Everyone in the room can point at the same screen. What you don\u2019t get is speed or scale across a book of business. To get things done, you need to know where to click, and you need to do those clicks yourself. That\u2019s the method that will continue to decline in popularity.<\/p>\n
2. AI Inside The Interface<\/p>\n
Google\u2019s version is Ads Advisor<\/a>. In April, Google announced it would gain policy troubleshooting<\/a>, around-the-clock security monitoring, and advertiser identity and business operations verifications. It can even make minor changes for you so you don\u2019t have to know which buttons to push in the UI.<\/p>\n In tools like Optmyzr, we have Sidekick, which operates tools on your behalf and can set up complex automations that previously required a lot of clicks in the UI. (Disclosure: Optmyzr is my company.)<\/p>\n Both are safe by construction. An assistant living inside an interface can\u2019t hallucinate the buttons it can click. Its changes land in the change log as expected. What it can do (and potentially mess up) is bounded by the buttons that already exist.<\/p>\n 3. AI Outside The Interface (The Agentic Shift)<\/p>\n The newest place to work is with your favorite agent, be it Claude, ChatGPT, Gemini, or one of the many others. By connecting the agent with connectors and MCPs, you give them a path to access data from other tools and enable them to push the buttons (albeit in a programmatic way).<\/p>\n Google publishes an official MCP server for the Google Ads API<\/a>, but it is read-only. It exposes just three tools: listing accounts, running GAQL queries, and describing resources. It cannot modify bids or pause campaigns.<\/p>\n Google\u2019s own spec sheet is worth reading closely. It lists the mode as read-only, then adds a parenthetical: current release. Writes are coming. They just aren\u2019t here yet, which means our industry gets the write path from everybody else first.<\/p>\n Third-party connectors are a different animal. Ours carries full GAQL as a read-only escape hatch, plus GA4 behavioral data, competitor overlap, vertical benchmarks, and account change history. It can read and write data.<\/p>\n Start with what each one can see. Google\u2019s server only reaches Google Ads data. A third-party connector can bring in margins, competitor data, on-site behavior, and the email where the client said to stop bidding on that product line.<\/p>\n So, let\u2019s take a closer look at what happens the day you stop managing Google Ads in the UI and shift to working with agents and MCPs to get through a day of managing PPC<\/a>.<\/p>\n 4 Things That Break When Moving PPC Outside The UI We usually think of hallucination<\/a> as the AI getting facts wrong.<\/p>\n MCPs introduce a different problem. A large language model is calling APIs that expect data a certain way, so its hallucinations can be about shape.<\/p>\n These APIs demand strict syntax. A small mistake about shape doesn\u2019t produce a small error. It produces a broken batch.<\/p>\n If an agent confuses a campaign for an ad group, it might creatively build dozens of new campaigns when you simply asked for a few new ad groups.<\/p>\n When it\u2019s reading, that\u2019s annoying. When it\u2019s writing, that\u2019s expensive.<\/p>\n 2. You Lose Track Of What\u2019s Where<\/p>\n In an interface, state is visible by default. What I mean is that when you add a keyword, you see it appear in the keywords table. If you need to verify that it\u2019s still there two weeks later, you go back to that table which serves as your source of record. In a chat, state lives in a transcript that scrolls away. So to know what keyword you added, you need to somehow find the right chat, and the place in the transcript that is relevant.<\/p>\n You can\u2019t let your entire business state live inside a scrolling chat transcript.<\/p>\n 3. You Lose Track Of What\u2019s Running<\/p>\n The scarier category is the automation you set up and forgot: scheduled agents, standing routines, and tasks that outlive the conversation that created them.<\/p>\n For example, you might set up a Claude scheduled task to check for money-losing search terms. It uses MCPs for reading the ads data and a separate one for sending a follow-up email.<\/p>\n If the email connector breaks, the automation fails silently.<\/p>\n These partial failures are often harder to detect than a total crash.<\/p>\n Silent failure is the default failure mode, and I know exactly what it costs.<\/p>\n Years ago I wrote a script to track account-level Quality Score. It worked perfectly on a small account. On a large one, it hit the 30-minute execution limit and died without an error report.<\/p>\n The weekly email simply stopped arriving, and nobody notices an email that doesn\u2019t show up.<\/p>\n By the time the client asked where their report had gone, the data was gone for good. Quality Score was a point-in-time attribute back then, and Google didn\u2019t offer historical Quality Score columns until 2017. If you didn\u2019t snapshot it yourself, it no longer existed anywhere.<\/p>\n You can re-run a report. You can\u2019t re-run yesterday.<\/p>\n If these risks feel new, they\u2019re not. We\u2019ve solved them the last time a major innovation came along in PPC automation<\/a>, specifically when Google Ads scripts<\/a> became popular.<\/p>\n What The Scripts Era Taught Us<\/p>\n Scripts arrived as a limited release in June 2012<\/a> and reached every advertiser globally that September. What happened next happened in a lot of agencies I\u2019ve talked to since.<\/p>\n Scripts proliferated. People copied them out of blog posts and pasted them in half-understood. Nobody wrote down what each one did.<\/p>\n Then the person who added it left the company.<\/p>\n It ran unsupervised for two years. Or it stopped silently, and nobody noticed for six months.<\/p>\n Eventually, somebody asks the question every PPC<\/a> lead has asked at least once in their career: Wait, what is changing these bids?<\/p>\n The governance did arrive. Script libraries, version control, failure alerting, a registry of what\u2019s running where.<\/p>\n It arrived years after the capability did, and every bit of the damage lived in that gap.<\/p>\n A broken script does the wrong thing consistently, which at least makes it findable. A hallucinating agent does the wrong thing creatively.<\/p>\n Ask yourself which one you\u2019d rather go hunting for.<\/p>\n How To Move Outside The UI Without Getting Burned<\/p>\n Most advice about agentic PPC stops at \u201ctry it.\u201d Here is what actually works, in the order I\u2019d do it.<\/p>\n 1. Start Read-Only<\/p>\n Connect the agent. Don\u2019t give it write access yet. You can easily set this up in the connector settings.<\/p>\n Spend a few weeks using it purely for reporting, analysis, and pulling data you\u2019d otherwise export by hand.<\/p>\n You\u2019re not being careful for its own sake. You\u2019re finding out where this particular agent gets confused, while being wrong still costs you nothing.<\/p>\n Google\u2019s read-only server is a reasonable place to run that experiment. Since Google\u2019s MCP is a bit difficult to set up, a third-party MCP for accessing your Google Ads data is a good alternative, and they carry the same controls.<\/p>\n 2. Demand Connectors That Are Honest About What They Can\u2019t Do<\/p>\n Nobody is vetting these connectors for you. Anthropic says as much on the screen where you add one: it \u201cdoes not control which tools developers make available and cannot verify that they will work as intended or that they won\u2019t change.\u201d<\/p>\n So check yourself. Before you trust a connector, ask the agent to list its own capabilities.<\/p>\n Good connectors ship a directory of what they expose and what they refuse. If it can\u2019t tell you where its limits are, that\u2019s your answer.<\/p>\n Pay attention to the write functions specifically. How many are there? What\u2019s the largest change a single call can make?<\/p>\n Watch for the overeager agent, too. It can go off the rails when it sees a capability but doesn\u2019t know the shortest path to a result. At one point, ours tried to build entirely new rule engine strategies just to fetch simple data, because it didn\u2019t realize a more direct, cheaper path was already available.<\/p>\n Nothing broke. But it burned time and tokens. When an agent proposes something elaborate, ask why. Often the answer is that it didn\u2019t know about the simpler option.<\/p>\n 3. Put Guardrails At The Connector Layer, Not The Prompt Layer<\/p>\n \u201cDon\u2019t spend more than $500\u201d in a system prompt is a suggestion, not a control. So is \u201cdon\u2019t touch the enterprise accounts.\u201d<\/p>\n Constrain by account instead. Connect one account first, and add more when you\u2019ve seen the agent behave. Most MCPs scope their permissions to what your username is allowed to do on the underlying platform. So, if you need a tighter boundary, set up a new login with fewer accounts connected or more restrictive permissions. Then there is no chance the agent reaches beyond those accounts if it starts to hallucinate.<\/p>\n The same principle applies inside the connector. Our current rule builders are suggest-only to ensure safety, and we\u2019re building toward a write path that lets the agent apply changes directly, provided the guardrails are enforced at the connector level. The raw query escape hatch is read-only, SELECT-only, and row-capped.<\/p>\n Those aren\u2019t instructions a model may choose to follow. They\u2019re enforced where the write would otherwise happen, which is the only place enforcement counts.<\/p>\n Blast radius should be a property of your setup, not a sentence you hope the model remembers.<\/p>\n 4. Make Memory Outlive The Conversation<\/p>\n Account context, client rules, and past decisions need somewhere durable to live.<\/p>\n We use a tiered approach: instructions at the Optmyzr account level, the specific ad account level, and the individual user\u2019s preference level.<\/p>\n Tools like CLAUDE.md files, at both the LLM and project levels, keep critical logic from disappearing when memory gets compacted.<\/p>\n The test for whether this is working is simple. If you have to re-explain to the agent every morning, your context isn\u2019t persisting. It\u2019s like onboarding a new junior every day and paying for it in tokens instead of salary.<\/p>\n 5. Ask For The Plan Before The Change<\/p>\n Preview everything before you apply it.<\/p>\n A prompt worth stealing:<\/p>\n \u201cShow me what you would change and why. Include the account, the entity, the current value, the proposed value, and the data that justifies it. Do not apply anything yet.\u201d<\/p>\n You get something you can review. You also capture the agent\u2019s reasoning while it still exists, before the conversation scrolls away.<\/p>\n 6. Keep A Decision Log, Separate From The Change Log<\/p>\n Google Ads change history is capped at 30 days<\/a> and doesn\u2019t even cover Editor changes. More importantly, it records what changed, and that the API did it under your email address. It records nothing about why.<\/p>\n![\"Google's](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/09\/02-google-ads-mcp-read-only-specifications-952.png\")
Google\u2019s MCP for Ads key specifications, listing \u201cMode: Read-only (current release).\u201d (Screenshot by author, August 2026)<\/p>\n
\n1. It Makes Things Up, Including Structure<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/09\/03-claude-scheduled-tasks-registry-list-908.jpg\")
Screenshot by author, August 2026
\n4. You Lose Track Of What\u2019s Broken<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/09\/04-claude-connector-permissions-89.jpg\")
Screenshot by author, August 2026<\/p>\n![\"Claude](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/09\/05-claude-project-memory-instructions-settings-385.jpg\")
Where to add a CLAUDE.md and other memory and instructions that your agent will always use. This shows where to manage these settings for a Claude project. (Screenshot by author, August 2026)<\/p>\n
![\"\"](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/09\/06-daily-brief-agent-mcp-status-626.jpg\")
Fred\u2019s vibe-coded daily brief shows when the agents and MCPs last did the thing they were supposed to do. (Screenshot by author, August 2026)<\/p>\n