The question my agent wasn’t asking.

I'm Jeremy Hurst, VP of Autonomous GTM at Swan. I left a $450K VP Sales role to take a job with a made-up title. I did it because I got a glimpse of the 100x seller - it stopped being a thought experiment and became the actual playbook. This is a field journal for GTM and growth leaders who can feel the model changing under their feet and prefer changing with it to being left behind.

I gave an agent 566 closed-lost deals from 2025 and told it to find new pipeline. The first version got it half right.

Half right wasn't bad - about $50K of pipeline per week, off 1 or 2 accounts. Then I changed one thing.

Same agent. Same pool of accounts. $250K in seven days. Here's what happened in between.

V1 was logical. For every closed-lost account, the agent pulled the CRM history, read the call transcripts, checked product analytics, worked out exactly why it died, & built a re-engagement play around it.

Solid. It should've worked better than it did.

But because every account started from a blank page, the 10th account got zero benefit from the 9 before it. All that judgment, gone the moment the sequence went out. Multiply that by 566 and you get a system that's slow by design.

I sat with why it was crawling. The answer was one missing step: categorization.

The agent was treating 566 losses as 566 unique problems. They're not. Closed-lost deals cluster into a handful of patterns. Without a step that names the pattern, the system can never learn from itself.

So I added it. Now, before the agent drafts anything, it has to land on a category. A repeatable pattern behind the loss.

Match an existing category & it pulls the plays that already worked. No match & it creates a new category, defines the strategy, & saves the account as the first worked example.

As an example, one deal died because the trial launched at the wrong time - in parallel to a massive fundraise. Aka, no bandwidth for a real test. That became its own category.

The next account that died the same way didn't start from scratch. It dropped straight into a shape it already understood.

566 accounts, one investigation each, zero leverage. That was V1.

V2 compounds. Every account either reinforces a pattern or creates a new one. The library gets richer & the work gets faster, at the same time. By account fifty, most losses match something already solved. The rare new pattern is the only thing that needs real thought.

That's the whole game - I'm realizing that thinking in systems instead of attacking things one off is the foundation of agentic GTM. You're not building a faster way to do manual work. You're building an asset that gets smarter with each rep.

$250K in seven days didn't come from more sends or better messaging. It came from a system that compounds intelligence.

getswan.com

THE SIGNAL

One question the agent wasn't asking, what pattern is this loss, was worth 5x the pipeline, because naming the shape of a problem is what lets a system learn from itself.

THE PLAY

The library build

• Pull your closed-lost deals from the last 90 days
• Before you touch outreach, ask the one question: what pattern is this loss
• Group the answers into 3 to 5 categories that keep repeating
• Build one play per category instead of one per account
• Test it on the next deal that closes lost this week, then let every match after it skip straight to a working play (roughly 30 minutes to set up)

SIT WITH THIS

The $50K to $250K gap came down to one step: teaching the system to remember. What does your agent forget every time it starts over?

COMMUNITY NOTES

Read more.

That one question is the only thing I do differently now.

-Jeremy