An Interview Should Show How We’ll Work Together

Episode four of 3 AImigo on hiring engineers who work with AI: work trials, assessing decisions, responsibility and choosing the next role.

When I was preparing to move into Big Tech, I was already working as an engineer and architect. Yet I struggled with algorithm problems far more than I wanted to. It was an uncomfortable discovery. At work, I could understand a problem and carry a solution through to a result. Put me in front of a timed exercise, and I felt I had to prove all over again that I could program at all.

I set aside six months to prepare. I spent several hours a day on algorithms and system design, weekends included. Every day, I recorded audio notes about what had worked and how I felt. Those recordings eventually became a nine-hour podcast episode. I had spent almost 180 days learning the separate skill of passing interviews, even though I already knew how to do the work.

I chose Google, Lyft and several European companies, and successfully went through the hiring processes at several of them. The nerves stayed with me. When you may have to wait a long time for another attempt, you want to reduce the part played by chance. I trained for the conditions of that particular exam. A system design interview felt like a timed chess game: your next move depended on where the interviewer took the conversation.

Figure 1 · Chess clocks and the working environment

Interview skill

~6 months of personal preparation

Work within company constraints

Tools, access, security, verification

The same engineer · different rules and actions

My six months of preparation helped me pass interviews. Working with agents under a company’s actual constraints requires other skills.

There is now another gap between that exam and everyday work. A company tests an engineer’s knowledge of Java. The engineer then joins a team, gets access to models, and starts solving problems with one or more agents. Their results depend on how they explain the problem, arrange verification, use permitted tools and meet security requirements. The interview may never have reached any of this.

In the fourth episode of 3 AImigo, I discussed hiring and careers with Alexander Polomodov and Evgeny Sergeev. Alexander had built hiring at scale in a large technology company; Evgeny brought his experience at Flo Health and in startups. The conversation was most useful where we disagreed. I thought preparing a practical interview with an agent could be fairly straightforward. Alexander argued that a good exercise tells you little about whether hundreds of interviewers can assess it consistently.

A company wants to know what work it can trust a person to do. A candidate is choosing the people and conditions in which they will do it. If an interview only helps someone perform the role of a successful candidate, both sides will have to start getting to know each other again after the person joins.

We discussed experienced engineers’ careers, leaving the first steps into the profession for the next episode.

3 AImigo, episode four, 4 September 2026. Watch the recording on YouTube

How a Company Tries to Know You Before You Join

There is a practical reason for a hiring process with several stages. Alexander recalled conducting a two-hour interview in 2016 alongside a platform specialist. His colleague assessed the candidate’s knowledge of the language. Alexander discussed their experience and asked them to design a small part of a system in a domain they knew. In one meeting, two people could gather enough information for a particular team.

That becomes harder when a company hires thousands of engineers. Few people can assess the technical details, judge how a candidate works with others, and help them understand why the job might appeal to them. Those people also have their own work to do. Companies therefore split the process into stages and train interviewers for each one, trying to maintain a comparable standard across teams.

Evgeny described a familiar combination of behavioural interviews, system design and a practical coding assessment. The work differs even within that combination. At Flo, a prospective manager is asked to clarify constraints, choose a suitable solution and organise the team needed to deliver it. An engineer is questioned in more detail about the proposed system. With a take-home assignment, there is a meaningful difference between a repository link that nobody can run and a solution with clear deployment instructions, checks and an explanation of the problem it solves.

Smaller companies often combine these conversations into one meeting. When I tried separating my own hiring process into several rounds, HR colleagues rightly reminded me that candidates drop out of a process that goes on too long. An extra stage has a cost before we even count the interviewer’s hours.

Figure 2 · Interview evidence and delayed feedback

Observations before hiring

Conversation · design · practical solution

Expectation of working together

Part of future work remains unknown.

Delayed evidence from actual work

Onboarding and internal processes delay feedback.

A convincing answer reveals only part of a candidate’s experience

  • Hire a weak candidate on the strength of a convincing answer
  • Lose a strong candidate because of the format
Alexander links multi-stage hiring to delayed feedback: the company tries to reduce uncertainty before someone joins, while risking the loss of strong candidates.

In Alexander’s explanation, a large company pays that cost partly because it takes so long to learn whether it hired the right person. A new employee must get to grips with internal systems and understand how the company reviews and deploys code. By the end of probation, they may still not have shipped a substantial change. A manager then struggles to distinguish a lack of ability from the way the organisation itself works. The longer a hiring mistake takes to become visible, the stronger the desire to reduce its likelihood up front.

An intensive selection process also turns away some good engineers. Some refuse to go through a lengthy procedure. Some get nervous. Others need more time for a difficult problem they would solve in a working environment. Algorithms and data structures still tell us something about a person’s preparation, but speed in answering a question cannot automatically stand in for speed in doing useful work.

AI has made the origin of the answer harder to judge, too. A model can produce code, prepare an architectural explanation and tailor behavioural answers to a company’s cultural principles. Confident reasoning used to suggest, at least roughly, that the candidate understood the subject. An interviewer now has to find out which decisions the candidate made themselves and what they understand in the explanation they present.

The Exercise Should Reveal How the Work Gets Done

At a small startup where Evgeny was CTO, candidates were given a paid working day with a real item from the team’s backlog. The company observed them, while they met the code and the people who might become their colleagues.

The startup had to prepare for that day. Its environment needed to run on a newcomer’s machine in about an hour; otherwise the time would disappear into untangling somebody else’s infrastructure. Evgeny recalled using Vagrant to put that environment together. What I like about this example is that the assessment went both ways. The team showed the candidate how it worked and took the risk that the candidate might dislike it.

Figure 3 · A paid day of working together
The team observes the candidate’s workThe candidate sees the team and process
Real backlog task
Ready environment~1 hour to start, then working together
Paid day

A case from Evgeny’s small startup. Vagrant helped prepare the environment.

In Evgeny’s startup, a paid day enabled mutual assessment: the team saw the candidate’s work, and the candidate saw the code and people. The environment had to be prepared beforehand.

Evgeny proposed taking the same logic into interviews with agents. Candidates once wrote algorithms with a marker at a whiteboard; later they were allowed to use a familiar development environment. Now we could allow the environment in which they work with AI and observe them solving a problem. Their opening moves provide material for a discussion: how did they explain the problem to the agent, which information did they consider necessary, and how do they intend to check the result?

His example was deliberately simple. A model proposes microservices and a distributed cluster for an internal back-office application with three users. I would ask the candidate to explain which need of those three users each additional service addresses, and whether that need justifies the expense and complexity. Such a discussion makes it easier to understand why the person accepted the model’s proposed architecture.

Figure 4 · The candidate’s working process
1Problem framing
Problem and constraints before work starts
2Context
What the candidate gives the agent
3Tool choice
Why this approach is chosen
4Result verification
What evidence is used to judge the result
Evgeny proposes observing how a candidate supplies context, chooses tools and verifies the solution. A convincing model output still needs justification.

I told a story about a friend in American Big Tech who repeated a joke at work about appending “make no mistakes” to a prompt. Someone treated it as a useful technique to adopt immediately. Before hiring, I would want to understand how the candidate intends to check an agent’s answers. Asking a model to make no mistakes tells me nothing about how a person will notice an error or what they will do next.

The same friend described his day of interviews at Apple, where he was back at a whiteboard, solving problems without his usual development environment. That format helps control the conditions of a particular meeting. It leaves the candidate’s work with an agent outside the assessment.

Alexander proposed distinguishing a person’s own engineering understanding from their ability to get a result with AI. For the first, we can discuss a data structure, an algorithm’s central idea or the reason a service failed, without turning the conversation into a race to type flawless code. For the second, we need to provide the permitted tools and observe what the person can accomplish with them.

It is easy to muddle these conditions. A remote interview can allow AI or restrict it; an in-person interview can do either as well. First decide what you need to learn to make a hiring decision. Then choose the setting, tools and means of enforcing the rules.

Figure 5 · Two independent choices: meeting format and AI use
Meeting formatPermitted tools
RemoteIn person
Without AI
With AI

What to assess: the candidate’s own understanding; results with tools.

First decide what to assess and which tools to allow

The meeting format and permitted tools can be chosen independently. Alexander proposes first deciding what to assess.

From this came the idea of extending the system design interview. A candidate could design a system, deploy it in an isolated environment, put it under load and investigate a failure. Running a service from their own design and testing it under load would let them explain the consequences of their architecture through the behaviour of a working system. A discussion about the failure could also reveal how they respond to disagreement and revise a decision.

Evgeny liked the possibility of bringing this together in a meeting lasting between 90 minutes and two hours, replacing some separate stages. He suggested starting with a small group of experienced interviewers who could later help train others. The format still needed to be tried, both to establish the cost of the meeting and to test how its results related to the candidate’s later work.

Figure 6 · A system on the test rig

Proposed interview stage · 90 minutes–2 hours

The design connects to a server. A pump applies load to the same server; a failure signal reaches the monitor, and diagnosis feeds back into the design.
  1. DesignThe design connects to the same running rig.
  2. DeployDeploy the system on the same rig.
  3. Apply loadThe pump applies load to this system.
  4. Diagnose failureFailure evidence informs a design change.
We discussed an interview in which a candidate deploys a system, puts it under load and investigates a failure. Its cost and scalability would still need to be tested.

A Good Exercise Still Needs Consistent Assessment

I proposed preparing a pool of ten to twenty practical exercises, with each candidate receiving one. For example, they might be given a service with an audit-log problem, an isolated environment, a terminal and internet access. The interviewer would know several reasonable ways to solve it and could observe the candidate choosing an approach, directing the agent and checking the changes.

For each exercise, I would want to define the conditions and several acceptable ways to solve it. In the conversation, I referred to BMad Method and OpenSpec as different ways of organising development with AI, including describing a desired change in a specification before starting. The candidate would need to choose an approach suited to the exercise and explain how it would help them carry the fix through to completion.

I had not built this pool when we spoke. My view was that the principles were stable enough to prepare exercises and describe several acceptable solutions. A candidate could also propose another well-founded approach. The interviewer would then have to understand why it worked.

Figure 7 · One task from a pool of 10–20

I propose preparing a task pool

One card is drawn from the task library. An arrow leads to a single task with a terminal and internet access.
Pool of 10–20 tasks

The library stays with interviewers.

One task for the candidate

An audit-log problem

Sandbox · terminal and internet

Different well-founded approaches are acceptable

I propose preparing a pool of 10–20 practical tasks and giving each candidate one. At the time of our conversation, the pool had not been built.

Alexander answered with a story about building hiring in a company with ten thousand engineers. He knew system design well, had written a book, and had prepared an assessment matrix and criteria. Yet new questions from interviewers arrived every week. A strong candidate could propose a correct solution absent from the handbook, and the engineer assessing it might not understand why it worked. The candidate’s expertise would then produce a poor score.

The reverse error was even more troubling. One candidate confidently solved a familiar problem in twenty minutes, overlooking a requirement for users to collaborate. Alexander brought him back to that requirement and asked him to revise the design. The candidate insisted that no change was needed. As Alexander worked through it step by step, the solution stopped holding together, and the candidate could not explain the changes it needed. Behind a well-rehearsed answer, Alexander did not find the expected understanding of networking, storage and computation.

A recruiter later told Alexander that the candidate had received offers from other large companies. That did not change Alexander’s own decision to reject him. He thought a less attentive interview could easily have accepted the practised answer. This was why he doubted that a new format could be extended to hiring at scale without substantial effort.

Figure 8 · Calibrating interviewers’ assessments
Task requirements

Two different examples from Alexander: a valid alternative and a familiar answer that omits collaborative work.

Different but valid approach

Risk: underrate a solution absent from the handbook.

Familiar answer, missed requirement

Risk: accept a familiar design without understanding collaboration.

Interviewer training

Interviewers’ questions feed back into criteria and training.

Scalability remains disputed

After clarifying that each candidate would receive one task, we still disagreed about scalability. Alexander describes two assessment errors: underrating a valid alternative and accepting a familiar answer without understanding a key requirement.

Our disagreement initially contained a misunderstanding. Alexander heard a proposal to examine every candidate on the full range of methodologies. I meant a library for interviewers, from which each candidate would receive one exercise. After clarifying that, we continued to disagree about the effort required to make interviewers from different teams assess solutions consistently.

I still thought preparing a practical exercise was manageable. Alexander maintained that a large company would have to keep training interviewers, discussing alternative solutions and replacing exercises after they leaked. A startup can ask a prospective colleague to solve a problem in its own product. A company handling large numbers of candidates also needs interviewers from different teams to agree in their assessments of comparable solutions.

Once I have prepared an exercise, I still need to help the interviewer recognise a valid solution beyond the handbook and spot a memorised answer that stops working when the requirements change. Interviewers need practice in assessing unfamiliar solutions. Common questions and a scoring scale help establish shared criteria, as described in OPM’s guidance on structured interviews. My proposal would still need to test agreement between assessments and their relationship to candidates’ later work.

A Broader Role Requires Your Own Understanding

To assess a solution, we first need to understand what work the company intends to trust the person with. In Alexander’s stories, similar answers concealed very different levels of preparation. I encounter a related difficulty in my work on AI adoption: finished code does not always reveal what the engineer can explain and check for themselves. In one case, I meet a strong computer vision specialist who knows the domain deeply but barely uses agents to develop tools. Their expertise remains valuable. They still need to learn the team’s way of working with agents so that they can apply that knowledge at the pace the team needs.

In the other, someone entered the profession after ChatGPT appeared, calls themselves a developer or DevOps engineer, and can put something together quickly. When I start discussing observability—the ability to understand a service’s state from its data and behaviour—it turns out they have simply been following Claude’s suggestions. They expect the model to tell them everything they need, even though they cannot always recognise a question it has left out.

I need an engineer who can explain how a solution works and use AI deliberately. Understanding the principles lets them decide what to check; working with an agent expands the range of tasks they can take on. Without their own understanding, a person will struggle to assess even a convincing answer from a model.

Figure 9 · Engineering understanding, AI tools and verification
1Base: Engineering understanding
2Optical body: Working with AI
3Focus knob: Explanation and verification
Observability

A question you don’t know to ask

Domain expertise without the new tool
Following suggestions without solid understanding
In my examples, strong domain expertise and confident agent use each leave gaps on their own. I look for an engineer who understands the system and can explain and verify a solution.

The same independence is needed when setting a task for an agent. While it remains a vague thought in your head, it is hard to hand it to someone else. You need to establish the constraints, express your intent and notice where it has been misunderstood. In an interview, I would therefore pay closer attention to how the candidate formulates a task and clarifies their thinking after receiving an answer.

I am also interested in how a candidate handles change. Subscribing to several channels tells me little about what they do when a new way of working arrives and they no longer have the energy to investigate it. A more useful conversation concerns how they learn, where they find an interest in new problems, and how they recover. The tools will change several more times during their time at the company.

Shared abilities do not make every role the same. Evgeny suggested looking at the users and the kind of uncertainty involved. In a product team, an engineer may work deeply within one domain on a particular problem for an external user. A platform team serves people inside the company, whose needs vary much more widely. In research, a person needs to be comfortable proceeding when both the result and the way to obtain it are still unknown.

Figure 10 · Three different engineering workspaces

Product: a user’s problem

A specific external user’s problem

Platform: different internal needs

Speed · scale · reliability · hardware

Trade-offs and evolution

R&D: work with the unknown

Dashed routes: directions still to test

Shared focus on outcomes

Evgeny distinguishes user outcomes, the variety of work done by internal teams and an R&D engineer’s ability to work with uncertainty. Alexander adds responsibility for trade-offs and subsequent changes to the system.

Alexander also left open the question of whether an autonomous agent system could do an engineer’s work when that work consists of shipping features from predefined requirements within an established process. In discussing product development, he advised developing a deeper understanding of the business and domain, so that you can discuss a problem with users and an agent in language they understand.

Alexander examined the platform case more closely. One team needs speed, another needs scale, and a third needs exceptional reliability or a difficult hardware integration. An engineer must make trade-offs among those needs and take responsibility for what the solution will become as further changes arrive. An agent can propose several options, and the engineer must justify the choice.

Individual features can look sensible and still add up to a product that is hard to maintain. It is therefore useful to discover in an interview how much responsibility a person can already take on. The stages may have the same names, but different things in the answers will matter to product, platform and research teams.

When an engineer divides tasks among agents and checks their work, some of what they do begins to resemble a manager’s work. At the same time, a manager can tackle technical tasks they would once have given to another engineer. The more departments a decision affects, the more strategy and agreement with people whose interests differ will matter.

How Engineering Roles and Salary Conversations Are Changing

A friend sent Evgeny an Engineering Manager vacancy at Anthropic listing a salary of zero pounds. It looked more like an error than a deception. Out of curiosity, he applied and received a response. The salary turned out to be high, while the hiring process was unexpectedly familiar. Even a lab at the frontier of AI wanted someone with the expertise to solve a particular team’s problems.

In another story, from an Anthropic conference in London, Evgeny described how a new initiative developed. One engineer builds a prototype and improves it several times. When the company decides to invest in taking the product to production, two more engineers join.

Figure 11 · One engineer starts; two more join

1 engineer: prototype and iterations

Decision to invest in deliveryFrom one engineer to three

+2 engineers: bring it to production

In Evgeny’s account of Flo

Justify a replacement hire

What work still needs doing? Is a replacement role needed?

In Evgeny’s account of Anthropic, one engineer builds and iterates on the prototype. Two more join when the company decides to invest in taking it to production.

At Flo, Evgeny said, replacing an employee who left already required a new justification. The vacancy no longer reopened automatically, because some of the work might have changed or been automated.

A broader engineering role appeals to me. When I was looking for work, I repeatedly found small differences between my experience and the requirements of narrowly defined vacancies. I could solve similar problems, but the particular language or combination of technologies differed. Now that an agent can help someone work beyond their familiar tools, this kind of filtering may exclude a person who could be useful to the company.

Evgeny connected the title Member of Technical Staff with an engineer who takes responsibility for solving a problem in full. As implementation becomes more accessible, choosing the problem demands more attention. A team can build many features quickly without gaining enough user traffic to test every hypothesis. Prioritisation and finding a way to test an idea become part of the engineering work.

When we discussed salaries, I suggested that differences in the value people create would become more visible. I had seen offers of €3,000 and of €10,000–15,000. Even the higher amount could be small relative to the contribution of an autonomous engineer working with agents. Those are my observations of individual offers and a hypothesis about a widening gap.

Figure 12 · Value created and the price of labour

Value created

My hypothesis about differences in value

Autonomy and responsibility
Demand for the profile
Who companies seek
Supply of specialists
What experience is available in the market

Salary offer

I suggested that differences in the value engineers create would become more visible. Alexander challenges a direct link between revenue per employee and pay, bringing demand, supply and autonomy back into the discussion.

Alexander challenged the direct link between a person’s contribution to a company and their salary. High revenue per employee does not mean the company will pass a corresponding share to that employee. Supply and demand affect the price of labour. Knowledge of a particular language may become less distinguishing over time. The difference between candidates may then become more visible in their autonomy and in the responsibility a company is willing to give them.

Evgeny reminded us that markets differ, citing the UK, where high pay can also be found in investment work and some medical specialisms. A comparison with another profession depends on the country where a person works and the experience they offer. A general account of high salaries in IT is not enough to guide that choice.

In an interview, I would explain the uncertainty I had already worked through, the decisions I made and how the work ended. That gives a company a basis for discussing the scope of responsibility it can give me. I can build some of that evidence in a project of my own if my current role does not offer a suitable task.

Curiosity Can Be Tested on a Problem of Your Own

Evgeny recalled entering IT in the early 2000s because he was interested in technology. In his first job, he wrote ASP, worked on page layouts and the server side, and dealt with Windows Server, MS SQL and indexes. He looked for answers in thick books. As products grew, specialisms emerged: it became harder for one person to understand every necessary detail.

Agents make it possible to try a broader range of problems again. Evgeny gave the example of someone building a personal notebook around their own habits and later noticing people with similar needs. Once others become interested in the notebook, its author will have to find out how closely their habits and needs match their own.

Figure 13 · Revisiting and testing a shelved idea
Shelved idea
Personal prototype

Learn what is missing

Verify the result

Possible value for others

An error sends you back to the task

The agent works overnight; the person checks the result in the morning.

Evgeny values seeing a problem and learning what is missing; Alexander suggests revisiting shelved ideas.

Alexander recalled ideas that people put off for years because of work and family commitments. An initial test might once have required a team and funding. Now agents can help them try part of the idea in a few days. Of course, they may wake up to a service that has stopped working. They will still have to check what they built.

A small project like this can also help with a career decision. Instead of a general claim that you are willing to learn, you have a task in which you learned the tool you needed, tested an idea and can explain what happened. There is something substantive to discuss at your next interview.

The Search Starts Before the First Application

The experience I have gained helps show what I can offer an employer, but leaves my own choice open. Alexander would begin a search by asking what he wants from the next period of his life, which problems interest him and which location works. He would then decide on the stage of a company’s development and the ways of working with others he finds acceptable. These criteria help a candidate understand what a future employer can offer them before comparing a vacancy’s requirements with their CV.

Next, they can research companies using public information. Employees’ talks reveal what they work on and how they explain decisions. The founders’ history provides a basis for questions about culture. A collection of vacancies can suggest the initiatives in which a company is investing. An agent can gather this information and compare it with a candidate’s criteria; the person then chooses which options to explore more deeply.

Figure 14 · A map of my career search
My criteria

Values · location · company stage · problems

Public evidenceTalks · founders · vacancies
Suitable companies

The path I choose

Role requirementsPreparation planMy experience
Alexander starts with values and criteria, then researches companies and compares role requirements with his experience. Preparation gains a concrete target.

Evgeny proposed making a personal application for this, which he and a colleague jokingly called localhost:3000. In such a working list, you could retain your research findings and the reasons you want to get to know each company better. As the search progresses, the list narrows. For the employers you choose, you still need to learn how to enter their hiring process and how much preparation it requires.

Figure 15 · A personal dashboard for job search and feedback

Evgeny’s idea for a personal app

Why it fits
My criteria and the supporting evidence
What to find out next
Questions for the team, information gaps
Next action
Compare experience, prepare, make contact

Application

Truthful experience matched to the role; the selected company’s rules.

Professional network

Experience from working together and a conversation with a person.

Information received

The person updates their notes and next action.

Evgeny proposes a personal app for research and preparation; I recommend allocating dedicated time to the search. The app can collect information about companies and questions for employees; the person chooses the next action.

After choosing a direction, Alexander would compare its requirements with his experience and decide which gap he was prepared to close. He had rebuilt his own coding practice through daily exercises, focusing on weaker subjects. Sometimes he switched to SQL, which came more easily, to recover a sense of progress. That preparation served to restore a skill he wanted for himself.

When preparing for the next role, I would also include structured learning in engineering with AI, combining study with my own practice. I would apply what I learned to a task where I could explain what I delegated to the agent and how I checked its work.

I would set aside time specifically for a search, allowing several months. In the conversation, I suggested three to four months as a planning allowance. That period needs to include researching employers and preparing according to the chosen companies’ rules. I would treat the search as a job in its own right.

You also have to account for the market’s actual rules. An engineer’s abilities may extend beyond the keywords in a CV, yet automated screening can still compare the document against a vacancy’s narrow requirements. I would therefore describe real experience so that its relevance to the particular work is clear.

Professional connections offer another way to begin that conversation. Alexander spoke about people who knew him through working together and through his talks: part of the introduction had already happened before a meeting about a new role. Evgeny saw the same value in conferences and professional communities. Sometimes a candidate can move directly to a substantive discussion rather than going through a long selection process. Usually, there is an accumulated reputation behind that opportunity.

Without a direct contact, I would use LinkedIn and people I know to find someone who could tell me about the chosen company. Even a conversation with an employee helps test the assumptions drawn from public research. A referral makes sense when the person understands which part of your experience they are willing to vouch for. I would allow time for cold applications as well.

Evgeny described people who, after a rejection, requested their personal data related to hiring under the GDPR to learn more about how the company had assessed them. It is useful to distinguish the right of access to personal data from the right to erasure. Access may help a person obtain information about themselves. It does not promise a complete copy of every internal conversation, because other people’s rights must also be considered.

An Agent Can Find a Company; People Choose Each Other

I would like to give an agent the task of finding companies against these criteria. I see the next step in a more explicit exchange of information between the two sides. In the conversation, I suggested imagining a candidate’s digital representative. It knows the person’s professional experience and life circumstances, researches companies, talks to their agents and finds suitable options. Given enough context, a model could process more information than I could study myself and spot a match I might never have reached.

A technical means for agents to exchange information already exists, such as Agent2Agent. The quality of a career recommendation, however, depends on what both sides disclose and how accurately it reflects reality. I saw this kind of matching as a possible development in which candidates and companies openly agree to use their representatives.

Evgeny had no objection to research and preliminary screening. An agent can compare companies, prepare a CV and cover letter, and answer basic questions. Beyond that point, he wanted to speak to a person. A future colleague will have to make decisions under stress and discuss them with the team. Their digital representative offers little basis for judging how they themselves will behave in that conversation.

Figure 16 · From agent matching to human decisions

Proposal for explicit agent matching

Research and match

The person’s experience and the company’s disclosed information

Preliminary match
People make the decision here
Human conversation

Decision to work together

Two independent human decisions
Evgeny and Alexander support agent assistance with research and screening. They leave the decision to work together to people.

Alexander drew the same boundary from the candidate’s side. Even if the agents consider Evgeny’s experience a good fit for another company, Evgeny still has to decide whether he wants to leave Flo. I was proposing a more explicit, substantive matching process. My co-hosts wanted to speak to their prospective colleagues themselves.

What frustrates me about current practice is how much of the automation stays hidden. A candidate uses a model to personalise and send applications. The company sorts them automatically. Both sides continue to act as though the process is entirely manual. We get stories of recruiters replying like chatbots and of instructions inserted into CVs to change an evaluating model’s behaviour. These attempts, known as prompt injection, expose the risk of mixing candidate data with instructions for the system.

I am more interested in an open exchange of reliable information. Evgeny also reminded us that automated sorting predates today’s agents. When applications outnumber what recruiters can read, companies use scoring algorithms. Candidates look for ways to make their experience visible, including referrals from people they know. With AI, both sides process information faster and still need to get to know each other personally.

Your Next Role May Be in Your Current Company

Evgeny suggested considering a move within a large organisation. A manager might want to return to hands-on engineering, already know the business, and have gone a long time without writing code. With agents’ help, they could rebuild some skills in a familiar context and learn during working hours.

Such a move can involve a temporary loss of status or pay. It is worth assessing that alongside the experience the person will gain and the direction they want to take. An external search usually means learning about a new role, product and group of people at once. Changing teams internally keeps some of those conditions familiar.

Figure 17 · Same company, different role
Same company, different role

Management approvals

Possible temporary trade-off

Status or pay may temporarily decline.

Direct engineering work

New skills during work hours

Domain knowledge carries overCompany constraints remain
Explicit rules within the workflow

A possible direction: move some late approvals into explicit conditions and checks. An instruction alone does not guarantee enforcement.

Evgeny proposes an internal move where the person already knows the business. Alexander describes returning to engineering work and the constraints he encountered at his new workplace.

Alexander explained why, over the previous year, he had decided to spend more time on engineering problems and AI again. At a senior management level, much of his time went to committees and approvals. His interests were moving towards technology and directly creating a result. He changed his management focus to pursue that work. He encountered constraints at his new workplace as well.

He wanted to understand more about infrastructure, model training and how models process requests, as well as how the labs themselves build the environments in which agents work. Building his own tools around models had begun to feel like chasing rapidly changing capabilities. He connected his next step with investigating the foundations of applied work and automating his own processes.

When choosing a role like this, I want to understand which actions the company permits an engineer to take and who is responsible for compliance with its rules. Some things previously checked through late approvals can be specified earlier and checked during development. Evgeny connected this direction with instructions for agents. Anthropic’s official AI-native SDLC playbook distinguishes instructions that make compliance more likely from automated checks that enforce a rule. Writing down a request does not, by itself, guarantee that it will be followed.

My six months of preparation helped me pass the selection processes on the chosen companies’ terms. Today, I would still take those terms into account. Before starting to prepare, though, I would want to know which decisions I would be trusted to make after joining, which tools I would be allowed to use, and how my future colleagues check the changes that reach users.

For my part, I need to be ready to show a task in which I can explain my own choices and their consequences.

Sources and Further Reading

Keep reading

Hiring & juniors 9 min

Hiring and career path in the AI era: what changes for junior, middle, and senior

AI is already changing hiring: test thinking, work with an agent, and ownership, not boilerplate-writing speed.