Contents
In brief
A standard algorithm puzzle in an online interview is solved by a model in seconds, and an invisible teleprompter such as Ghost GPT hides that from the interviewer. Bans and proctoring lose this race. On Habr, the founder of SymbioWay describes the opposite move: allow the agent officially, and change the task so it helps only someone who can split the work, check the result, and answer for the code. That is closer to watching a real task than to an hour of screening that never proved much.
What happened
The author has assessed people on live tasks since 2008–2010 and founded the company in 2014. The claim is blunt. A tool does not replace knowing the work. A surgeon with a Da Vinci robot and a photographer with a modern camera still have to understand what is in front of them. The model removes routine. Only someone who understands the work can split it, write the instruction, check the answer, and carry the code.
The skills he treats as current are not "passing LeetCode." Look wider at neighboring areas: a Java server developer who has seen the client, the build, tests, and mobile. Hold architecture, distributed systems, algorithms, and networks more deeply. Know how to run the tool: the spec, context economy, subagents, several agents, token spend. Stand closer to the product and the customer, because a person is legally responsible for the code.
He has always considered a one-hour screen unreliable, and a full trial day too expensive. With an agent, watching someone work on a real slice of a project got cheaper. The company hands out multi-level examples as a file through a Telegram bot. The market note in his text: IT vacancies in Russia roughly halved from mid-2024 to mid-2026. Hiring did not get easier. It got different.
Why it matters
If the task tests something a model already does better than a person in a minute, you are measuring access to a teleprompter, not an engineer. A ban turns the race into hide-and-seek. Permission without a new task measures the same thing, only honestly, and still misses.
A useful signal shows up where the agent runs into understanding. The candidate has to notice a wrong result, narrow the context, and keep out of the model anything that must not be sent. You can see that in architecture and in a live slice of a system. You cannot see it in "invert a tree."
In practice
- Do not build a screen on standard algorithms a model closes immediately. If that round stays, do not treat it as proof of level.
- Allow the agent in writing and say what may be used. A hidden teleprompter and an allowed agent leave different traces. The second can be read from the conversation and the diff.
- Shape the task so that without a system sketch and a check of the result, the agent produces plausible, useless code. The author's multi-level examples live in the bot's file, not in one function.
- Watch how the person splits the work, what they put in context, and what they throw out. Saving context is as much a skill as choosing a data structure.
- Replace an hour of talk with a short observation of work on a slice of your project or a teaching project. A full trial day is still expensive. A slice is not.
- Leave responsibility with the person. The model does not sign the code and does not attend the incident review.
Takeaway
An agent in an interview does not retire the engineer. It retires tasks an engineer can no longer tell apart from generation. Change the task, not the proctor's rulebook.
Comments