AI Now Writes Much of the Code. Here Is How Inventrium Vets the Engineers Who Direct It

By Oladipo Ajayi, CTO, Inventrium

A traditional coding interview answered one question: can this person turn a problem into working code under pressure? That used to be most of the job. It no longer is.

This week, Felix Rieseberg, an engineering manager at Anthropic, wrote that he has stopped using traditional coding interviews for his team. Short coding challenges have become a weak signal, he argued, because AI tools can now complete them well.

The big platforms are moving the same way. Meta has been developing AI-enabled coding interviews, and Google has piloted a coding round that allows an approved AI assistant.

The question for hiring managers is not whether to drop the coding test. It is what the job has become, and how to test for that.

The work has moved from writing code to directing it

Watch how experienced engineers now build with AI and the shift is plain. At AI Engineer Europe this year, Matt Pocock walked an audience through his workflow from a vague client request to a shipped feature. Very little of it was typing code. The human work sits at both ends of the process:

Clarify the request. Before any code is written, the AI questions the engineer until every open decision is settled: what earns points, what happens to existing data, what is out of scope.
Write down the destination. The agreed decisions go into a short requirements document: the problem, the user stories, the implementation and testing decisions, and what will not be built.
Slice the work. The plan becomes small tickets with clear dependencies. Each is a thin slice that works end to end, so integration problems show up early, not at the end.
Let the agent build, tests first. The AI picks up a ticket, writes a failing test, writes the code to pass it, and runs the tests and type checks before each commit.
Try it and review it. People test the feature and review the code. In Pocock's demo, every automated test passed, yet using the feature exposed a missing database table.

Two lessons follow for anyone hiring engineers.

First, the plan decides the output. A vague brief produces confident, wrong code, faster than ever. The engineer who can turn an unclear request into a sharp plan is now worth more than the one who types fastest.

Second, this is not "write a spec and stop looking at the code." Pocock says he tried that approach and it failed. Engineers still need to understand the system and shape it, especially the boundaries between its parts. What has changed is where their effort goes: less typing, more deciding, structuring and checking.

A syntax test measures none of that. In our market, where many strong engineers are self-taught and many CVs now read the same, a test that AI can pass tells you little about the person. Korn Ferry's 2026 talent acquisition research points the same way: 73% of talent leaders ranked critical thinking as their top hiring priority, with AI-related skills fifth.

What we now assess

For engineering roles, we assess how well a candidate directs and verifies the work, not how fast they can type it. We look for evidence of six things.

1. Turning a vague request into a clear plan

We give a deliberately incomplete brief. Strong candidates ask the questions that settle the open decisions, then write a short plan an AI agent could work from: the goal, the user stories, what is out of scope, and how we will know it works.

2. Systems thinking and architecture

Can they slice the work into thin pieces that each work end to end, and spot what depends on what? Can they design clear boundaries between the parts of a system, so it stays easy to test and change, for people and for agents?

3. User empathy

Technical work should make life easier for somebody. We look for candidates who think about the customer on a slow connection in Ibadan, the finance officer who is not technical, or the support team that has to explain an error when something breaks.

4. Design thinking

"Working" and "good" are not the same thing. We ask candidates to explain their choices: what they prioritised, what they left out on purpose, and what they would improve with more time or evidence.

5. Directing and verifying AI

Can they give the AI the right context and constraints, insist on tests before code, and catch what it gets wrong? We look at what they checked, what they changed, and what they refused to trust. Passing tests are evidence, not proof.

6. Depth of analysis and clear communication

We ask candidates to explain a trade-off in plain language. If someone cannot explain why they chose an approach, they may not yet understand its risks well enough to own it.

How our vetting works

Inventrium does not send employers a pile of CVs. People who do the job set the test for their own field, and experienced outside assessors help with senior roles. Every shortlisted candidate passes four stages:

Scored screening. Skills, experience, salary expectation and availability, plus a review of the CV, GitHub and portfolio. We give real weight to what candidates have built on their own.
A plan-first practical test. For engineering roles, candidates get a realistic, deliberately vague brief. They write the plan, use AI tools to build from it, then review and test what comes out. We score the plan, the slicing and the review at least as heavily as the finished code. Live coding is used where a role needs it.
A technical interview. Mostly about "why": candidates defend their plan and their choices, and tell us what they would change with more time, less budget or ten times the users.
Final checks. Salary, notice period, availability and why they want this role. Reference, background and education checks are available too, always with the candidate's consent.

We agree the scoring criteria with the hiring manager before the search begins, and more than one assessor reviews the evidence where it matters.

What hiring managers should change

Start with the outcome, not the list of technologies. Ask:

What must this person achieve in their first 90 days?
What business problem will they own?
What decisions will they make without close supervision?
What evidence would show that they can do the work?

Then test the work as it is done today. Give candidates a vague brief, ask for the plan an AI agent would follow, let them build with the tools they would use at work, and ask them to review what came out.

The bar has not dropped. Someone still has to understand the system well enough to shape it and to catch what the tests miss. That is the skill to hire for. We are not lowering the bar. We are moving it to where the work now is.

Hiring tech talent?

Since 2021, Inventrium has found, tested and placed tech and digital staff for employers across Nigeria, and for remote teams hiring African talent. We recruit across software engineering, IT, data and AI, design, product, cloud, cybersecurity and digital roles.

You receive a small, tested shortlist, usually 3 to 5 candidates per role within 3 to 5 weeks, each with a scored report covering strengths, risks, salary expectation and availability.

Request a talent search. Looking for your next role? Join our talent network on the same page. We never charge candidates a fee.

Sources

Published by Inventrium Magazine