Welcome to another episode of "Things That Make Doug Wonder If He'll Have a Job Next Year!" Today we're diving into the world of recursive AI agents and how to connect them to your internal data sources. If you missed the livestream, don't worry, we've got you covered with this breakdown of Jeremy Ross's Clay wizardry and Jordan's barely-contained anxiety at watching someone with 17 million Clay credits flex on all of us.
The Future Vision: Cannonball 3.0
Before we get into the technical weeds, let's paint a picture of where we're headed. Jordan kicked things off by outlining his vision for how go-to-market systems will evolve in the next 6-9 months. The future is a triangle, with, wait for it FOUR key components:
1. Systems of Record: Your CRM, customer database, product data in Snowflake, aka, the ground truth about who your customers are. This is information you already have but is likely a mess (duplicate information, leads and contacts not properly mapped).
2. Systems of Information: This is where go-to-market leaders are under-investing. Think of it as what replaces ZoomInfo (which can’t be soon enough): accurate ground truth about your market and prospects, not just your customers. Owner.com's system of information would include data like which restaurants are on DoorDash, their review counts, and star ratings.
3. Systems of Action: Orchestration layers like Clay that tie into tools that execute actions (LinkedIn automation via Heyreach, hashed ad IDs for individual targeting on Facebook, Nooks or Aurum for phone dialing).
At the center of it all: The System of Intelligence – a model like Claude or ChatGPT that maintains context about your intentions and can understand the gap between what you say you want and what you actually need.
The Gong Integration: Connecting AI to Your Call Data
Jeremy shared a fascinating demonstration of how he's leveraging Clay to supercharge Gong data analysis. The real magic here is turning transactional data analysis into ongoing, contextual intelligence that gets better over time.
Here's what he built:
A system that pulls Gong call transcripts and metadata
A method to maintain conversational context between runs using OpenAI's responses API
A way to feed each subsequent call analysis into the same "chat in the cloud"
The business impact? As Jeremy explained: "It saves managers from having to listen to the whole call. You have much more focused one-on-ones that are straight to the point... AI does a great job of really honing in on some of the subtleties in these conversations."
Even more powerful: the system tracks individual reps over time, identifying trends across dozens of calls. It can also analyze team-wide performance, uncovering what's working, what's not, and what the enablement team should focus on.
Recursive AI Agents: Self-Improving Prompts
Now for the really mind-bending part: Jeremy demonstrated how to build AI agents that iteratively improve themselves through recursive feedback loops.
Here's how it works:
Start with a prompt that instructs the AI to perform a specific task (in Jeremy's example, researching a LinkedIn profile)
The agent runs and produces an output
Another agent analyzes that output against the original prompt to evaluate performance
If the evaluation score isn't perfect, the system automatically rewrites the prompt
A new row is created in the table with the improved prompt
The process repeats until the output hits a perfect score
Jeremy's system even uses conditional logic to prevent infinite loops (though as Jordan pointed out, there's still a risk of bankrupting yourself through recursion if you're not careful—the system can create up to 50,000 rows).
Why This Matters: The Intent/Reality Gap
So why go through all this trouble? As Jordan explained, it's about bridging the gap between our intentions and reality:
"The purpose of recursion is for the model to capture your intent as well as possible, and then to bring that to actual subject matter experts and say, 'I know that I said this, but now that I see it actually run on the real world, I didn't really mean that.'"
When you build AI agents in isolation, they operate in a vacuum without real-world data. The recursion loop allows the AI to learn from actual data and improve its understanding of your intention.
Jordan found that about three recursions is the sweet spot where the model becomes "happy with itself" and produces high-quality results.
How to Start Building Your Own Recursive Agents
So how do you take steps toward building recursive AI agents? Here's the practical advice distilled from the conversation:
Start with the APIs you have access to: Clay, OpenAI, Gong, etc.
Use ChatGPT or Claude to help you: As Jeremy noted, "When I first started using the responses API, I just copied the docs from OpenAI, dropped that in ChatGPT and said, help me figure this out."
Begin with a simple recursive pattern:
Initial prompt → Run → Evaluate output
If evaluation < 100%, rewrite prompt → Run again
Repeat until satisfaction
Use Clay's conditional run feature to prevent infinite loops (e.g., "Only run if score < 100")
Connect to real data sources like Gong to give your agents context
The key point: You don't need to be a developer to do this. With enough persistence and the right tools, you can push the boundaries of what's possible today.
The Future Is Coming Faster Than You Think
As Jordan summed up: "The people that push the edge of the tools are those that will be able to reap the rewards faster. But eventually, this is going to become dramatically democratized."
What Jeremy is doing manually today will likely be built into products within the next three months. The question is: will you be ready to use these capabilities when they arrive?
"There are no right answers," Jordan concluded. "There's just better attempts."
Helpful Resources
If you're looking to dive deeper into recursive AI agents, check out these resources from our content library:
The Cannonball GTM Prompt Library and Guide: April 2025 Edition - Start with the right prompts
Growth Metrics Accelerator: Using AI to Optimize Your Targeting & Reduce CAC - Learn how to measure the impact
Inside the Machine: How We Built an Automated PVP Generator with Clay, Claude, and Gemini - See another example of Clay+AI integration
Have questions about building recursive AI agents? Join us at our weekly office hours every Friday at 11am pst to discuss further.
Share this post