Agents Are the New SaaS: A 7-Day Playbook for Building AI-Native Businesses
The central thesis is that SaaS sold tools; agents sell completed work. Where traditional SaaS told a team "here is software to help you," an agent product says "here is a job your team no longer has to do." Because the underlying market is human labor — a multi-trillion dollar opportunity — the host contends the TAM dwarfs anything the SaaS era unlocked.
The host offers a clear filter for finding viable agent workflows: the job must happen frequently (hourly beats daily), have a measurable finish line, touch software that already exists, involve edge cases that are learnable but not pure human judgment, and cause pain the buyer can feel financially. The sweet spot sits between basic automation (which Zapier can handle) and open-ended judgment calls (which early AI breaks on).
Before writing a line of code, founders are told to shadow a human doing the job — screen recordings, narration, 10–20 real examples — then write a formal agent spec covering trigger, context, tools, autonomous actions, approval gates, escalation rules, and success criteria. The recommended first build is deliberately minimal: a "draft and approve" or triage agent, not a fully autonomous system. Trust is built through visible logs, controls, and handoff rules, not raw capability.
Validation follows a concrete sequence: - Run the workflow manually with Claude or ChatGPT before building software - Create an eval set from 50 real examples with marked correct answers - Sell two pilots in the same niche at fixed or outcome-based pricing (e.g., $1,500 setup + $1,000/month, or $30 per qualified appointment) - Productize only the parts that repeat reliably across customers
The bottom line is a 7-day sprint: pick a niche (home services, property management, insurance agencies), interview operators on day two, spec the agent by day four, run it manually on day five, ship the smallest useful version by day six, build an eval set on day seven, and spend weeks two through four selling pilots and publishing workflow teardowns to drive distribution. The host is direct that the window is open but that most people are watching the shift rather than building into it.
The host argues that AI agents represent a generational business opportunity analogous to the SaaS wave, but bigger — because the addressable market is human labor itself. The episode lays out a detailed framework for identifying, building, and selling agent-first products to SMB niches in under a month.
I handle this one annoying job better than a junior employee, faster than an agency, and it's cheaper than adding headcount.
The agent does the work but the wrapper creates the trust. Customers actually need to see what happened. They need to see logs, approvals, controls, handoff rules.
If every roofer needs the same emergency call script, service area check, financing question, and estimate follow-up, boom, you have a product.
you want to be in the business of selling painkillers, not vitamins
Find the smallest painful workflow that repeats all day in a niche that you understand and make it disappear — answering a phone, booking a job, triaging the ticket, updating the system, escalating that weird case
Building agents is the new SAS.
We saw billions of dollars of value creation during the SAS era.
The total addressable market for agents is just way bigger. You know it's human capital.
Labor is a multi-trillion dollar market.
daily is good but hourly is better
If a workflow is too basic, basically the problem is basic automation you know zaps and stuff like that can do it. If it's pure human judgment, the first version will break. So the sweet spot is repetitive work with enough judgment that AI can help.
People are buying agents for the first time ever. So they don't want all of it at the same time, especially if you're not Microsoft or you're not Salesforce.
your eval set is basically like the gym. Every time you change the prompt, the model, the tools, the workflow, the agent goes back through the gym and is able to basically know what's good and know what's bad.
outcome pricing is the future of how a lot of these agent first businesses and software is going to be priced
By the end of week four I have formats that are working, I know where to double down, I know where I can spend paid money to acquire customers
agents are the new SaaS because software is moving from help me do the work to do the work with me
People charge thousands of dollars for this type of thing, but it is free like always on the Startup Ideas podcast
A lot of people are just not partaking in the shift — they see the shift, they understand the shift, but they're not building agent-first businesses
SAS sells software, agent SAS sells work. A normal SAS product says here is a tool a team could use. But an agent SAS product will say here is a job your team no longer has to do by hand.
A good agent workflow has five traits: it happens all the time (daily is good but hourly is better), it has a clear finish line, it touches already existing software, the edge cases are annoying but learnable, and the buyer can feel the loss.
When speccing out your agent, it should have seven key parts: what wakes the agent up, what context does it need, what tools can it use, what is it allowed to do itself, where does it need approval, when should it escalate and bring a human in the loop, and what does success look like.
There are four good first versions of a minimal useful agent: a draft and approve agent, a triage agent, a coordinator agent, and a bounded action agent.
The fastest path is usually a pilot where you manually do the work with AI and then you productize the repeated parts. Start with three customers in one niche — same niche, same workflow, same pain — and sell the outcome.
Pricing examples: $1,500 setup and $1,000 a month for one workflow; or $2,000 setup plus $30 per qualified appointment; or $3,000 a month up to 500 handled tickets.
Distribution: think about teardowns, think about poking fun of the old way, think about creating memes around it, and then creating content, picking the winners, putting paid ads against it
Day one: pick a niche where missed work costs money — home services, property management, insurance agencies
find the job, shadow it, spec it, run it manually, build the smallest useful agent, sell the pilot, then productize the repeatable parts
Pick one niche and write down 20 jobs people complain about. Score each job on five things: how often does it happen, how expensive is the pain, how easy is it to know when the job is done, what tools does it need access to, and who already owns the budget.
Once you find the job, before you start coding, shadow a human being who does the job. Watch someone do the job 10 to 20 jobs. Ask them to screen record it. Ask them to narrate what they're doing.
Take 50 real examples of the job and mark the right answers, then run the agent system against them to evaluate whether it classified the problem correctly, asked for the right missing information, and used the right policy.
we tested this on 50 of your old maintenance requests. It routed 42 correctly, flagged six of them for human review, and made two mistakes. Here are the two mistakes, and here's how we fix them.
Show the old way of doing a process and then show the agent way. Pick one workflow, make the internet associate you with it.
make the checklist, make the benchmark, make the tear down, make like 50 examples of this workflow post, and you're going to be in the content game
Use some of those assets that are starting to work and then put paid ads around it
Focus on one platform to start
Day two: interview 10 operators, ask them to screen share the workflow, keep the calls as research, and just watch them
Day three: pick one workflow with frequency, pain, software access, and a clear success metric
Day four: write the agent spec — trigger, context, tools, rules, handoffs, eval
Day five: run it manually with AI — use Claude or ChatGPT and copy and paste the context, draft the output, ask the human to approve — you are testing whether the AI helps before you build the software
Day six: build the smallest useful version — draft and approve or triage is usually enough here
Day seven: create the eval set from 50 real examples
Week two: sell two pilots in the same niche
Week three: add the product wrapper — the logs, the approvals, the settings, the analytics, the handoffs — and use AI to actually build that software
Week four: publishing workflow teardowns, turning the pilots into proof, and doubling down on content strategy
Shared with Earmark · earmark-ai.com