How to vet a dev agency before you sign

The decision

You've decided an agency will build your MVP, you have a shortlist, and the pitches all sounded good — that's their job. Between those pitch meetings and a signature sits the real decision: which of these teams can actually deliver, and how would you know? You are buying something you cannot inspect, from people whose profession includes being convincing. When this goes wrong, it typically burns 50–100% of the original budget plus 4–8 months — so the hour of diligence per agency described below is the cheapest insurance you will ever buy.

The questions that actually determine it

These are questions you answer about each agency, not questions to ask in a meeting.

Who will actually write your code — and did you meet them?

The single most common agency failure is the bait-and-switch: seniors run the pitch, juniors do the delivery. Ask to meet the specific people who will build your product — not the founders, the builders — and ask for their names in the contract. A good agency does this without friction. Refusal, or vague talk of "resource allocation," tells you everything.

Have they built something whose shape matches yours?

"We've done lots of apps" is not evidence. What matters is shape: same domain, similar integrations (payments, logistics, whatever your product touches), similar scale. Pick the portfolio project closest to yours and ask forensic questions: what was hardest, what would they do differently, who on the current team worked on it. Rehearsed answers stay general. Real experience gets specific fast.

What do their references say when you ask about the second project?

Everyone's first project goes fine — that's the honeymoon, and it's what curated references are selected for. Ask each reference three things: what changed after the first release, what happened when something went wrong, and whether they went back for more work. References that all trail off after project one are a pattern, not a coincidence.

How do they handle change — in the contract, not the conversation?

Your spec will change; every MVP's does. Scope-change pricing is where fixed bids become open wallets. Read the change-request clause before you sign: how changes are priced, who approves them, what happens to the timeline. A fixed bid with no change-request clause isn't generosity — it's a fight scheduled for month two.

What do you get if you walk away mid-project?

Assume the engagement might end early and read the contract from that angle. Does the code live in a repository you own from day one? Is IP assigned to you as work proceeds, not on final payment? Can another team pick up where they left off? These clauses decide whether leaving is possible at all, and they are only negotiable before you sign.

Your options, with honest costs and risks

The options here are levels of vetting depth, not builder types. Each level catches what the previous one can't.

The minimum bar

Reference calls with the second-project question, portfolio forensics on the closest-shaped project, and a meeting with the actual builders. Cost: a few hours per agency and some awkward conversations. This catches the sales-team-in-front-of-a-junior-shop pattern, which is the most common failure. What it can't catch: whether the code itself is any good.

The paid pilot

A scoped, real (not throwaway) task of one to two weeks before the main contract — a small feature, an integration, a technical prototype. Cost: ₹80,000–3 lakh ($1,000–4,000) and two weeks of calendar. What it shows you is how they actually work: communication rhythm, how they handle ambiguity, what their delivered work looks like, whether estimates hold. It converts the biggest purchase of your year from a promise into a sample.

The independent technical review

An outside senior engineer — no stake in the deal — reads the agency's proposal, their public or sample code, and the pilot output if there is one. Cost: ₹40,000–1.5 lakh ($500–2,000). This is the only level that can judge the code and architecture on your behalf, and it turns "the pitch felt solid" into an informed opinion. It also quietly changes the agency's behavior: teams quote differently when they know an engineer will read the proposal.

What I'd recommend

Nobody should skip the minimum bar. Not for a trusted referral, not under deadline pressure. It costs hours, and it catches the failure mode that costs the most. When I've engaged agencies myself, the second-project reference question did more work than everything else combined.

If your committed spend is above roughly ₹10 lakh ($12,000), run a paid pilot. The two-week delay feels expensive when you're eager to start; it is nothing against the 4–8 months a failed engagement costs. If the spend is above roughly ₹15 lakh ($18,000), add the independent review — at ₹40,000–1.5 lakh ($500–2,000), it pays for itself the first time it flags an inflated estimate or a shaky architecture.

And regardless of budget, walk away immediately — no negotiation, no second meeting — on any of these:

• They refuse to name or introduce the people who will write your code.
• A fixed-bid proposal has no change-request clause.
• The code will live in their repository, not yours.
• Every reference relationship ends after the first project.

Each one predicts the same ending. The pitch quality of the agency in front of you has no bearing on any of them.

When this doesn't apply

• You already have a trusted technical person — a committed co-founder, an advisor, a senior engineer friend who'll spend real hours on this. Have them run the technical side of the vetting; this guide is the substitute for that person, not an addition.
• The "agency" is one known freelancer with a brand. Then you're vetting an individual — track record and bus factor — not a team, and the meet-the-builders and staffing questions don't apply.
• Regulated domains — health, finance, anything with a compliance regime. The vetting above is necessary but not sufficient; you also need someone who can judge domain-specific competence, and that requires a specialist.
• The engagement has already started and is going wrong. That's a rescue, not a vetting problem, and it needs hands-on human help quickly.