Founder Firesides

Eddie Kim on why agents shouldn't start from a blank canvas

Eddie Kim· Co-founder & Head of Technology at Gusto at Gusto
·~32 min·English·Y Combinator
AgentsAI CompanyBusiness Strategy
TL;DR

Gusto's co-founder built an AI business partner for small businesses by starting from the recurring tasks they already do — not an open-ended prompt — routing exact jobs like payroll to deterministic crons, and shipping the whole thing with five people in ten weeks.

01Core Mental Model

The Blank Canvas Problem

An all-powerful, open-ended agent paralyzes a small-business owner — so Gusto Co-founder starts from the tasks they already repeat every week, not an empty prompt.

We actually start with all of the things that Gusto is already solving for our customers.

— Eddie Kim, Founder Firesides
Key Insight
The scarce resource for a non-technical owner isn't capability — it's imagination about what to automate. Gusto's edge is that it already knows the answer, so it can suggest instead of ask.
02How He Learned It

You Don't Get It Until You Build It

Reading about agents and building one yourself are completely different experiences — Eddie spent eight hours setting up an air-gapped agent, and that hands-on experience shaped Co-founder's product and technical design.

you really don't get it until you actually set it up yourself and you actually have hands-on experience.

— Eddie Kim, Founder Firesides
Key Insight
The product's best features — texting the agent, the 30-minute heartbeat — weren't dreamed up in a strategy deck. They were copied from what surprised Eddie when he lived with an agent himself. Hands-on use was the real market research.
03Product Design

Chat Is the Interface

The engineer's instinct is to build a richer UI; the winning move is to make the agent smart enough that a text message is enough.

everybody is blown away by the ability to run everything they're doing on Gusto through text message

— Eddie Kim, Founder Firesides
Key Insight
Every hour spent polishing a client is an hour the user still has to spend learning it. Pushing capability into the agent instead of the interface is what lets a payroll run collapse into an SMS reply — the least intimidating surface a non-technical owner already uses.
04Architecture

Heartbeat or Cron?

A heartbeat — an LLM running every 30 minutes — is flexible but probabilistic, so anything that must be exact, like payroll, gets a deterministic cron instead.

there's some things you actually want to be deterministic and so like payroll probably one of those things

— Eddie Kim, Founder Firesides
Key Insight
The open-source heartbeat pattern is seductive because it's simple, but 'the model will probably do it around the right time' is unacceptable for money movement. Co-founder's real engineering was detecting which trigger a task needs — and not paying to run an LLM every 30 minutes when a plain cron would do.
05Why Gusto Wins

Your Own Data Makes It Proactive

Fusing each customer's own Gusto history with aggregate industry data lets Co-founder suggest automations proactively — including surfacing a $50,000 R&D tax credit one business didn't know it was owed.

we actually found them $50,000 in R&D tax credit. They didn't really know that that was possible.

— Eddie Kim, Founder Firesides
Key Insight
The prototype that just built web apps was a demo; the version that mattered fused the prompt with Gusto's per-customer and per-industry data. That fusion is the defensible part — a competitor can copy the chat box but not the years of payroll and HR history behind it.
06Go-to-Market

All Upside, No Resistance

Selling automation to a small business faces none of the job-loss fear enterprises do — the owner is the one drowning in the busywork, so removing it is an easy sell.

They just want to be able to do more with less.

— Eddie Kim, Founder Firesides
Key Insight
The demo sells itself because the buyer feels the pain personally — the hour every week spent moving data from MindBody to a spreadsheet to Gusto before payroll even starts. In the enterprise, the person you're automating is rarely the person signing the check; in a small business they're the same person.
07How They Built It

Working Code Was the Only Spec

Five people shipped a tier-one launch in ten weeks by deleting process — no meetings, no specs, no Figmas — and letting throwaway pull requests replace documents.

We didn't have meetings. We didn't have any text specs. We didn't have any Figmas.

— Eddie Kim, Founder Firesides
Key Insight
When writing code is nearly free, the cheapest way to make a product decision is to build all the permutations and look at them — a working PR carries information a spec never can. The discipline shifts from planning what to build to being willing to throw most of it away.
08The Macro Bet

A Step-Function Easier

Starting a business got gradually easier for fifteen years; AI is a step-function jump on that same curve — and the census data already shows more people taking the leap.

more and more people are seeing this as a viable path for them to pursue

— Eddie Kim, Founder Firesides
Key Insight
Gusto's original bet — making the back office easier — and its AI bet are the same trend at two scales. If automating payroll, compliance, and HR removes the last reasons not to start, the total addressable market isn't just today's small businesses; it's everyone who almost started one.