Your Agents Are Only as Whole as Your Data — Part 2: Wholeness over Fragmentation
July 22, 2026
Chris Carolan, Trisha Merriam and Erin Wiggers test Wholeness over Fragmentation against what agents run on: your data. Trisha's HubSpot-to-Salesforce sync died silently when four account records met one; Erin demos Emergent, one context layer instead of ten. Self-test: count your tabs.
Moments from this episode
Key takeaways
Gaps and disagreements between your systems do not stop an agent — they invite it to fill the hole with an assumption, which is where hallucination starts.
A fragmented data model can fail completely silently: Trisha's HubSpot-to-Salesforce sync died because four account records on one side met one on the other, no case was created, and nobody was notified.
The opposite failure is just as common — an agent that refuses to move on something it could do, because a best practice says the data looks wrong, when the humans around it know the discrepancy is intentional.
Count your tabs: how many do you have to open to answer one question about your own data? Two is fine, three to five is annoying, more than five means you have the problem.
Even when flipping between ten tabs costs you nothing mentally, it still takes longer than having the answer in one place, and you pay that cost every single time.
Store the invariants as facts — this client renamed deals to visits, these two records are one person and must never be merged — and inject them into every agent run instead of trusting a search to surface them.
Rank your knowledge by how far an agent may stray from it: facts are never deviated from, canonical documents are superseded only deliberately, and everything else is color.
More context is not automatically better context; a million-token window still behaves like cognitive load, and an agent asked to hold an entire book before doing a task performs worse, not better.
For a document too big for any window, fan it out: slice it, give each agent one slice, and pass the synthesizing agent the summaries and facts rather than all the raw pages.
Hunt for know-how that lives only in someone's head — when a person says their process is three steps and you watch them open ten applications, a paper binder and a phone call, you have found it.
Moving data between systems is cheap once the rules are written down, so mismatched property names are no longer a reason to email a PDF between teams.
The single-source-of-truth rule was a storage constraint, not a user need — decide from the user's perspective rather than IT's, and put the data where the work is happening.
Transcript
This transcript was generated by AI and may contain errors. It is provided to help you skim and search the conversation.
Keep learning
You just watched the work. Now build it.
The next step is doing it yourself — live, with people on the same path.
Get the next conversation in your inbox
One signal-dense note when a new episode lands. No noise.