We Turned Our SaaS into a CLI and Skill...... Is This the Future of SaaS?

We recently rebuilt Lessie AI around a different assumption: the primary user is no longer a person clicking through a dashboard, but an agent executing tasks.

Lessie started as a pretty standard SaaS for people discovery and outbound. Search, filter, inspect, export, send emails. It worked, but the interaction model started to feel dated.

The shift isn’t just “AI inside SaaS.” Software is becoming an execution layer for agents.

So we stopped asking how to add AI, and started asking what the product looks like if it’s designed for agents first.

APIs aren’t the abstraction

We thought clean APIs would be enough. They weren’t.

Agents don’t just need endpoints. They need guidance on how to use them:

what to do first, when to stop, when to ask for clarification, and how to recover from failure.

That logic now lives in what we call a skill layer. It’s less like API docs, more like a runbook.

The stack changed less than the behavior

On the surface, we added two entry points: a CLI and an MCP-accessible skill.

But the bigger shift wasn’t interface, it was behavior.

Most of the “product intelligence” moved into this layer that tells the agent:

when to use Lessie, how to choose execution mode, when a query is ambiguous, and when not to proceed.

Letting agents use your product breaks things

Once we actually let agents run workflows end-to-end, a few patterns showed up quickly:

• they don’t reliably infer environment state → mode selection has to be explicit

• they don’t handle ambiguity well → disambiguation needs to be enforced

• they retry broken paths → fallback logic has to be built in

And most importantly, APIs don’t encode judgment. They return data, but not whether the result is useful or worth continuing.

The workflow is the product now

A request like: "Find 20 CTOs at AI companies in SF and verify their ML background" used to take 1–2 hours across tools.

Now an agent can run the search, review edge cases, filter weak matches, and return structured output in minutes.

The difference isn’t just speed. The product is no longer the UI flow, it’s the set of actions that can be executed reliably.

Where we landed

We don’t think SaaS disappears. But the center of gravity is shifting.

The UI becomes just one client. The durable value moves to:

how well an agent can operate your product without supervision.

So the questions we care about now are simpler:

Can an agent decide when not to act?

Can it recover from failure?

Can it avoid wasting resources?

Can it complete the workflow end-to-end?

Curious how others here are thinking about this. Are you still building AI features into SaaS, or starting to design around agents as the user?