Last week I was trying to figure out what should be automated in GTM and when a human has to stay in the loop. This week I tried new tools, ran even more parallel workflows, and found myself overwhelmed by the hidden cost of all of it: someone still has to manage and review the work.
The week in one minute
What we’re pondering: Models and harnesses are merging, and supervising too many agents fries my brain.
What we’re trying: Direct mail to cut through the noise, Pipedash on our own deals, and GTM Codenames.
Worth your attention: Mythos 5.1, Salesforce inside Claude, AI coworkers with real identities, and why Clay cannot define your ICP.
GTME jobs: 933 open roles. Three interesting ones below.
Stranger than fiction: Fake Costco employees made almost 8,000 AI videos and got 107 million views.
What we’re pondering
1. The model and the harness are merging
Dan McAteer wrote on Latent Space that the model is not the product. The harness is: the loop, tools, context, memory, permissions, and the thing you actually sit in. But the more interesting thing happening now is that the harness and model are merging. Claude Code and Opus already feel like one choice. Codex means OpenAI models. Cursor was the big independent harness, and after the SpaceX deal, OpenAI said it intends to wind down models in Cursor by November 12.
We are starting to choose a model and a harness as one thing as harnesses feel more like highly personalized tailoring than hoodies you can throw over any model. As I keep ping-ponging between harnesses, one thing that keeps me sane is knowing the Mada Brain is helping preserve and transfer context between them, and making a daunting task like trying my chief of staff system in Grok Bot take minutes.
More on my Grok Bot impressions in a future issue. In the meantime, Angela Ferrante is co-hosting an amazing Grok Bot builder event today with the SpaceXAI team at Amplitude HQ.
2. The brain fry is in the supervision
Yesterday I found myself trying to count my concurrent workflows. I had 12 between Codex, Grok Bot, and a few things still in Cursor. Every few minutes I switched, prompted, reviewed, gave feedback, then tried to remember what the first workflow was doing. I felt like a manager with 10+ direct reports in the early days of Branch, except the reports are agents and they never end. By the end of the day I felt fried. The hidden cost of AI is not just tokens. It is the toll that comes with constant context switching. BCG and UC Riverside researchers call it “AI brain fry.” Until the GTM “dark factory” becomes reality, our only recourse is to limit the number of active projects, just like we limit the number of direct reports a manager can handle well.
What we’re trying
Direct mail, and a book you can flip
Email and LinkedIn DMs barely work now. So much of it is written by bots, including mine, and even a pretty good outbound message can disappear into the pile and it’s assumed to be AI-written and delivered.
So Dan Ahmadi and I are trying to break the pattern by getting attention with actual physical things. He is doing the Cryptex. I made How to Attract a Dragon, an attribution book you can actually flip through. Now you can flip through it online, but we’re also making physical versions. Comment or reply if you’d like us to send you one!
Pipedash on our own deals
We ran Pipedash on three recent closed-won deals. We built the system to give 0% credit to the hand raise, and to actually credit what led to the demo request. The things that got us those deals were previously untrackable before: content we made with Carilu Dietric, champions who moved jobs, and people talking about Upside because they love it. Those are the things that usually disappear from attribution because they are hard to measure.
More here: Running our Pipedash attribution model on our own GTM data.
GTM Codenames, and a game night
AI makes it much easier to make fun, slightly unnecessary things. Last weekend Dan and I decided GTM needed its own Codenames. We came up with many of the words ourselves, then used AI to help design it, and now we are printing it. We playtested it and it was harder and more interesting than regular Codenames. Come play with us at GTM Codenames Game Night.
Worth your attention
Claude Fable 5.1 and Mythos 5.1 are out today. They are still the same model, but with different levels of safeguards. I haven’t tried it yet, but excited to explore it.
Salesforce and Anthropic announced Claudeforce, Salesforce inside Claude. Pilots are live, with an open beta planned for September. I am curious to see its adoption and how it differs from the existing Salesforce MCP, which Dan found has significant limitations. I personally agree with Ben Fuller ’s point: Claudeforce will only be successful with net-new Salesforce customers because Salesforce doesn’t even use Salesforce internally the way it recommends its customers should.
Optimizely launched Virtual Teammates, AI coworkers with their own identities, persistent memory, permissions, and recurring work. This is a much bigger shift than another prompt box, and the marketing org chart is starting to include software workers. Excited to have Alexander Atzberger on our podcast soon and explore!
Matt Heinz argues AI may be making your best people slower: the tools create output quickly, but someone senior still has to read, fact-check, and fix it. This is exactly the brain-fry tax I felt this week.
Loren Shumate made the point I keep running into: knowing Clay does not mean you know the customer. AI can scale a good GTM strategy, but it can also make a bad one move much faster and waste dollars.
If you want GTM in your ears: Future of Marketing with Camille Ricketts , Ethan Smith , and me, plus Pavilion’s Revenue Leadership Podcast.
GTME jobs
The live internal GTME tracker had 933 open roles. Outbound was still the biggest category, and workflow automation appeared in 21.7% of roles.
Credal, Forward Deployed Software Engineer, Brooklyn, $140K to $220K: enterprise AI-agent deployment.
Gooseworks (YC W23),Founding Creative Engineer, San Francisco or remote, $120K to $200K: distribution and self-improving ad creative. Love this one <3
NOX METALS, GTM Growth Engineer, Los Angeles or Detroit, $70K to $140K: GTM engineering for AI-powered manufacturing.
Stranger than fiction
Fake Costco “employees” made close to 8,000 AI videos and got at least 107 million views. The videos were eventually traced to one affiliate operation. The woman in the Costco uniform is not a real person, and she sends people to a gift-card survey that does not exist. The same play was run against Aldi, Target, Home Depot, and Walmart taking demand capture with no human in the loop to the next level.