Shipping the agent is the first half.Making it better is the second.

In this episode, Mickey Alon sits down with Sunil Subhedar, who led growth marketing at Anthropic after roles at Uber, Canva, and Google. They discuss why most agents stop improving after launch, and what it takes to build the loop that makes them better.

Almost every product team shipped an agent this year. Ask them how it has improved since launch and the room goes quiet. They tuned the prompt, added skills, added tools. Very few can tell you what actually changed, or how they know.

Sunil has run growth at consumer scale and inside one of the companies building the models, which makes him unusually well placed to see why B2B never built this muscle.

Below are the main takeaways from the conversation.

Agents are becoming the users of your UI

"People are not going to go into these UXs anymore. Agents are going to go into these UXs."

Sunil's point is that most teams are building for the world they can currently see. They take a heavy configuration surface, rebuild it inside their own product, and put an AI in front of it. Nothing about how the work gets done has changed. The setup, the rigor, the clicking, all still there, now with a chat box attached.

To be precise about what this does not mean: the workflows do not go away. The databases, permissions, and audit trails matter more, not less, because something is now executing against them on a user's behalf. What goes away is the assumption that a human will sit down and configure it by hand.

The consequence is uncomfortable for anyone mid-build. If your product's value is concentrated in a configuration surface, and you are still investing in teaching users to navigate it, you are investing in the layer that is disappearing. If a user has to spend three months learning your UI to reach their goal, they will find something that lets them state the goal instead.

The loop broke at the last step, for structural reasons

Build, measure, learn was supposed to be how software gets better. In consumer, volume forced teams to build that loop for real. B2B teams could build, but they measured badly, and applying what they learned meant a ticket, a sprint, and a conversation about the legacy codebase.

Sunil's explanation comes down to two structural facts. Your N is tiny, especially in enterprise, so there is barely any signal. And the signal that exists arrives six to eighteen months after the decision that produced it. That explains most of what B2B growth teams do instead: one or two variants rather than fifty, arguments settled by opinion, attribution stories assembled after the fact.

One of those two is changing. The lag is not going anywhere, but the signal per interaction is climbing fast. When someone talks to your product instead of clicking through it, you get their intent in their own words, the execution recorded, and the outcome observable, on every interaction. Clicks never gave you that.

What an agent actually needs in order to get better

You shipped an agent six months ago. What has to be in place for it to improve, and in what order? He gave five things, and the sequence is the useful part.

1. The data and governance layer first. Not exciting, but everything downstream inherits its problems. What data matters, and is it clean.
2. A goal, and a written definition of success. This is where he thinks most teams are quietly stuck, tuning a prompt or a skill without any way to know which dial to turn.
3. Log the run, not just the output. Every engagement, so you can compare what the agent did against what you hoped it would do.
4. Both feedback loops. The quantitative half starts cheap, as thumbs up and thumbs down. The qualitative half is the one that used to be impossible: not just was this bad, but why. What makes it usable is a rubric with real categories, so a failure tells you which component to go fix.
5. An agent pass, then a human, with the human shrinking. Let the agent improve against your guidelines, keep a person in the loop, and expect them to matter less as edge cases get covered.

Read that list again and notice what it has in common. Not one of those five was required in order to ship the agent. Every one is required for the agent to get better. That is precisely why so many teams have the first and none of the second.

Start from the goal, not the dashboard

Sunil's test for whether a team has any of this comes from his own habit as a growth marketer:

"I start with the measurement framework before I begin any sort of work. And I don't think that's a common thinking process."

Mickey's corollary is about where teams look. Most teams ship features, open a dashboard, and hunt for a signal. That is backwards, because a dashboard mixes so many signals that none of them are actionable. Start from the goal, work backwards to where people fall off before reaching it, then run something specific there, with the right thirty or hundred accounts rather than with everyone.

NPS is the cleanest example of the failure. Advocacy matters, but when did yours last move, and could you say why? You get a ten and a zero and no mechanism for understanding either.

Context is the constraint, not the prompt

A million lines of code is deterministic. A million words of prompt is not, and endlessly tweaking it is not a strategy. Sunil watches teams accumulate skills that clash and duplicate, instruction files that try to hold everything, and agents running slowly because every request drags archival material along. His mental model for the fix: a clean desk beside a well-stocked shelf. Bring down what this task needs, use it, put it back.

The fix is not a better prompt. It is assembling the right knowledge and actions for this user, this role, this goal, right now, and leaving the rest on the shelf. His related habit change: "Don't be very solution oriented. Be very problem oriented and goal oriented." Senior people are the worst offenders, because their instinct is to lead the witness.

The org shape

A central group has to settle governance, data access, and permissions once, so nobody hits the same bottleneck on their fourth project. Handing a team AI licenses without that plumbing is how adoption fails. Then three roles inside each team: an architect who pairs systems thinking with domain expertise and owns what great looks like, a solutions engineer who takes what works on one laptop and makes it work for hundreds, and a program manager who watches how people actually use it and carries that back.

His warning for small teams is that one person doing all three gets stuck. He is also skeptical of the chief AI officer touring each team to diagnose problems. What works is champions who already understand what their group needs.

Shipping an agent was the easy half

Most teams have done it. The half that decides who wins is whether anyone can see what the agent did, tell success from failure, and get that back into the system fast enough to matter. Three separate things, none of which shipping an agent required, and all of which are now the actual product.

The agentic interface layer is the right thing to be obsessive about because it is where the evidence lives. It's the only complete record of what people are actually trying to do with the product. No amount of dashboard work reconstructs that after the fact.

That points to one thing product leaders should stop doing. Stop asking your ICP to adapt to your product. Build a system that adapts to your ICP, measure whether it is doing that, and close the loop fast enough that the answer changes month to month rather than year to year. Build the brain, buy the operator.

The brain is the agents that automate your domain. The operator is the agent your users talk to inside your product, which is what the Foldspace Product Agent is.

Sunil Subhedar is an Entrepreneur in Residence at Storm Ventures. He previously led growth marketing at Anthropic, and worked at Canva, Uber, and Google. He writes the Models & Mortals newsletter.

Mickey Alon is Co-Founder and CEO of Foldspace, co-author of Mastering Product-Led Growth, and previously co-founded Gainsight PX and Insightera. He writes about the shift from click-driven software to agentic interfaces.