Start with a Minimum Viable Agent, not a moonshot

When a small firm decides to try an AI agent, the instinct is to aim big. If you are going to build something, the thinking goes, make it count: the agent that runs the whole sales pipeline, or handles the entire inbox, or manages every supplier. These are the projects that get sketched on a whiteboard with great enthusiasm and quietly abandoned three weeks later, because they were too big to finish, too complex to trust, and too entangled with everything else to test safely.

There is a better starting point, and it is deliberately humble. Dharmesh Shah, a software founder who builds agent tools, coined the idea of the minimum viable agent: the smallest version that does one useful job from start to finish. Not the grand system you can imagine, but the modest one you can actually stand up this month, watch work, and rely on. For a small firm with no AI department and no time to waste, the minimum viable agent is the only sensible way in.

Why moonshots fail in small firms

A big agent project fails for reasons that have nothing to do with how good the technology is. It fails because it touches too many things at once, so when one part misbehaves you cannot tell which part. It fails because it takes weeks to build before it does anything useful, and a busy owner runs out of patience and momentum long before then. And it fails because it is hard to trust: a system that does ten things is ten times harder to check than one that does one thing, and an agent you cannot check is an agent you cannot use.

Shah's case for agents is genuinely optimistic. He has written that the future of much software is agentic, software that takes steps toward a goal rather than waiting for each click. But optimism about where the technology is going is not a reason to start big. It is a reason to start, and starting means picking something small enough to finish.

What a Minimum Viable Agent is

The minimum viable agent does one repeatable job, end to end, and nothing else. It has a clear trigger (something happens), a short sequence of steps (it does a few things in order), and a clear output (a result you can check). That is the whole specification. If you cannot describe your intended agent in those three parts in a sentence or two, it is too big to be your first one.

The discipline is in the word minimum. You are not building the best possible version. You are building the smallest version that genuinely helps, getting it running on real work, and only then deciding whether to add to it. Most agents that earn their keep in a small firm stay small for a long time, because small is what makes them reliable.

Pick one repeatable job

The right first job is one you do often and in roughly the same way each time. Frequency matters because a small agent that saves you ten minutes on a daily task is worth far more than an ambitious one that saves an hour on something you do twice a year. Repeatability matters because an agent follows a pattern, and a job with a stable pattern is a job an agent can hold.

Good candidates tend to be unglamorous: turning enquiries into tidy records and a first-draft reply, assembling a standard report from the same few sources each week, preparing the brief you need before every meeting. Pick the one that is most repetitive and least risky, and leave the exciting, sprawling idea for when you have learned what these tools actually do well.

Build the smallest agent that genuinely helps, prove it on real work, then decide whether to grow it.

Test against a real problem

The most useful habit Shah recommends is also the simplest. Do not evaluate an agent, or any AI tool, in the abstract. Test it against "a specific problem you want to solve." A tool that demos beautifully on someone else's example tells you nothing about whether it helps with your actual work. So bring your real problem, your real data, your real edge cases, and see whether the agent handles them. If it does, you have something worth keeping. If it does not, you have saved yourself from buying capability you would never have used.

This keeps you honest in both directions. It stops you adopting a slick tool that does not fit, and it stops you dismissing a plain one that does. The only test that matters is whether it moves a real problem you have.

It is worth being concrete about what "small" buys you. A tightly scoped agent fails in obvious ways: when it goes wrong, it goes wrong on one job, and you can see exactly where. A sprawling one fails quietly, in the seam between two of its many tasks, and you can spend an afternoon working out which part misbehaved. Small is not a beginner's compromise you grow out of. It is the property that makes an agent trustworthy enough to leave running, which is the whole reason you built it. The firms that get value from agents tend to run several small ones with clear edges, not one large one nobody fully understands.

The honest limits

Keep the scope tiny, and keep it tiny on purpose. The pressure to add "just one more thing" to a working agent is constant, and every addition makes it a little harder to trust and a little harder to fix. Add only when the addition clearly earns its place, and check the whole thing still works when you do.

Review on a rhythm, too. Shah's own practice is to re-evaluate his tools roughly every quarter, because the field moves and what was the right choice in spring may be outclassed by autumn. A minimum viable agent is not a build-once-and-forget asset. It is a small, living tool you revisit: keep it if it still earns its place, retire it if something better has arrived, widen it only when the case is obvious. The smallness is what makes that easy.

What to do this week

Write down one job you do often that follows the same steps each time. Describe it as a trigger, a short sequence, and a checkable output. If it fits in a sentence or two, it is the right size. Build the smallest agent that does just that job, test it against a real example with real data, and run it for a fortnight before you even think about a second one. One small agent that works will teach you more about what is possible in your firm than any number of whiteboard moonshots.

Working out which job is the right first one, and keeping the scope honest so it actually ships, is the practical heart of the AI Advisory for Teams work. If you want a first agent that earns its place rather than a plan that never finishes, book a business call.

Sources and further reading

• Dharmesh Shah, "Why The Future Of AI Is Agents," Simple.ai, 25 September 2024. Shah's case that much software is becoming agentic, from someone who builds these tools for a living.
• Dharmesh Shah, "Feeling Behind on AI? How Dharmesh Shah Stays Up to Date," Agent.ai blog, 2025. His habit of testing any new tool against a real problem, and re-checking his stack every quarter.