Design the MVA

The cheapest thing in your company is now a feature. You describe one in a sentence and it works by lunch, which is exactly the trap dressed as a gift. So founders pile up panels and integrations and call it a product, while the one question that decides whether the company lives goes untested: will a real person let a machine own the job the whole thing depends on. Ship the smallest agent that proves that loop, not the feature pile that flatters you.

The method

Full method, the Goldilocks scoring filters, a worked MVA scope, anti-patterns, and a copyable template: references/mva-method.md.

Step 1 - Find the Goldilocks task

Pick one task that is high-volume, guideline-based, and low-risk, all three at once. High-volume, so the loop runs often enough to save real time and throw off real signal. Guideline-based, so there is a knowable right answer a human can check at a glance. Low-risk, so a wrong answer is an annoyance and not a harm. Your highest-value task is usually your highest-risk one. Write that on the roadmap for later. Earn trust on the boring task next to it first.

Step 2 - Define the Minimum Viable Agent

The Minimum Viable Agent is the smallest agent that proves the intelligence loop closes: task in, agent acts, outcome out, human accepts or corrects, the system records the choice. The deliverable is not a screen. It is a working loop around one task, and it can be thin on the surface: a text box, an email reply, a row someone approves. The intelligence is the product.

Step 3 - Name the one high-risk hypothesis it tests

State the trust bet so it can come back no: users in role R, facing task T, will accept the agent's output without redoing it at least M percent of the time. One hypothesis, not a multi-feature dashboard. If a proposed build does not map to a bet that can fail, it is scope creep, however reasonable it sounds.

Step 4 - Add the self-healing fallback

Decide in one bounded sentence what the agent does when it is unsure or the model fails: retry a set number of times, switch to a backup, then drop to a fixed non-AI path and escalate to a named human. A pilot user who hits a dead end twice goes back to doing it by hand. You lose them on reliability, not capability.

Output

• The MVA spec: the one Goldilocks task, the loop it proves, the single high-risk hypothesis, and the bounded fallback.
• Run measure-ai-native-pmf next, once the loop is live and the hypothesis has a clean yes or no.

Constraints

• One task, not ten. The unit of work is an experiment, not a feature.
• Pick the Goldilocks zone, not the hardest problem. The first wrong answer on a high-risk job detonates the trust your whole roadmap needs.
• The MVA must prove the loop or it is just a feature. Run the Remove-the-AI test: take the model out, and if the loop still works, you scoped a feature.
• Stay theme-agnostic; the founder supplies the domain, you supply the rigour.

Dictionary

Minimum Viable Agent · the Goldilocks zone · self-healing fallback · the Remove-the-AI test

Copy-paste version

For non-coders: paste into any chatbot or no-code builder (Claude.ai, ChatGPT, Lovable, Bolt, v0).

Act as my AI-native product coach. My idea: [ONE_LINER]. Who it's for: [ICP]. Domain: [DOMAIN].
Help me write my Minimum Viable Agent in one sentence, in this exact shape:
"My first agent does [the single task] for [ICP], proving that [the loop], and when it fails it [the fallback]."
Rules:
1. The task must pass all three Goldilocks tests: high-volume (runs often), guideline-based (a
   knowable right answer a human checks at a glance), and low-risk (a wrong answer is an annoyance,
   not a harm). If my chosen task is my highest-value but highest-risk one, push back and tell me to
   pick the safe job next to it instead.
2. The loop must prove a real human hands a real job to the machine and keeps the result. If removing
   the AI leaves it still working, tell me I scoped a feature, not an agent.
3. The fallback must be bounded: retry a set number of times, then drop to a fixed answer and a named
   human. No infinite loops.
Don't flatter me. End with the ONE next action to build it.