Most AI agents that fail don't fail at the build stage. They fail at the scoping stage — usually because the builder skipped it entirely. Here's the 5-question framework we walk every Squidgy builder through before Ace generates a single prompt.
1. Who specifically is the agent for?
Not "coaches." Not "small businesses." Specifically. "Strength coaches with 30–80 online clients who use TrueCoach for programming." That's a target. You can verify the fit, you can find them, you can talk to them.
If you can't describe the user in 12 words or less, you don't know them well enough yet. Talk to five before you build anything.
2. What specific job is it doing?
One job. Not three. The agents that ship are the ones that do one well-defined thing — not the ones that try to be a Swiss-army platform.
"Helps with content" is bad scope. "Drafts three Instagram posts per week in the user's voice based on their week's training logs" is buildable scope. The first version of an agent should fit on an index card.
3. What does success look like for one user, on day 30?
If a user pays and uses the agent for a month, what concrete outcome should they get? "They'll save time" isn't enough. "Their newsletter will go out weekly for four weeks straight, with a 40%+ open rate" is enough.
This becomes the metric you build to. It also becomes the only honest pitch you have when someone asks why they should pay.
4. What does the agent never do?
Scope is mostly defined by exclusions. A coaching agent that doesn't give nutrition advice is safer than one that does. A legal-firm intake agent that doesn't opine on legal strategy is far more deployable than one that does.
5. What's the minimum number of integrations?
Each integration is a chunk of complexity, a place where things break, and a piece of ongoing maintenance. Build with the absolute minimum on day one — usually one or two external systems, often just the user's calendar or CRM.
Add integrations only when a paying user complains they need it. Most agents need fewer integrations than their builder thinks.
The five-question test
- Who specifically is it for? (12 words or less, names the persona)
- What one specific job does it do? (one sentence, no "and")
- What measurable outcome does the user get on day 30?
- What does the agent explicitly never do?
- What's the minimum integration set on day one?
If you can't answer all five
Don't build. Talk to five users instead. The agents that fail at the build stage usually had a vague answer to question 2 or 3. The ones that ship had crisp answers to all five before any prompts were written.
On Squidgy, this scoping happens in conversation with Ace before any agent design starts. The clearer your answers, the faster Ace can build.