AI is changing how software is built. In some teams, it already runs the execution cycle.
AI developer and serial founder Aleks Bykhun, whose products reached 400K+ users across AI, NFTs, and developer tools, says engineers are beginning to delegate key decisions to AI agents. In this interview, he explains how this shift may reshape software development.
JOIN US ON TELEGRAM
Follow our coverage of the war on the @Kyivpost_official.
Q: You’ve worked through multiple major cycles in technology. What makes the current wave of AI agents different?
Aleks Bykhun: Earlier tools made me faster at tasks I already understood. Claude is different. It sometimes tells me my implementation is wrong. I review the reasoning, and it turns out to be correct.
That changes the dynamic. It stops being a coding assistant and starts functioning as a technical reviewer.
What unsettles me is that it evaluates decisions. When a system can analyze trade-offs and suggest a better solution, you reconsider how much of the process should remain manual.
Q: How would you describe what is changing in software development right now?
AB: I see six stages in how engineers hand responsibility to AI systems. In practice, it means higher output with less control.
Stage one was autocomplete tools like GitHub Copilot in 2021. They suggested code, but the developer still decided everything.
Stage two is agent-based coding. You paste a ticket into Claude Code, receive a draft solution, and review the pull request. Many engineers already accept changes with little review.
I’m somewhere between stages three and four. In stage three, AI reviews analytics, user feedback, and revenue data, then proposes what to build next.
Stage four reverses the loop. AI runs the implementation, and the human approves.
I already apply this in small ways. I ask the agent to question my assumptions before proceeding. My role moves from writing code to supervising systems that generate and test it.
Q: What happens when the industry reaches stage six?
AB: The possibility of reaching stage six is what pushed me to start a new company. At that stage, the primary users of many digital products are no longer people but AI agents.
Even now, agents process documentation and interact with APIs directly. They read structured content and execute tasks.
Today, an engineer might compare Supabase and Neon by reading documentation and testing free tiers.
Soon an AI coding agent can run that evaluation itself. It reviews the docs, makes API calls, checks error responses, and selects the tool that performs more reliably in automated tests.
If documentation is unclear or API outputs are inconsistent, the agent chooses your competitor. In that environment, machine-readable structure and predictable behavior matter more than design.
Q: You’ve described a future where AI agents choose devtools. Did you build 2027.dev to address that?
AB: Yes. For years, devtool teams optimized for human developers through polished onboarding and readable documentation.
But AI agents don’t act like humans. They parse documentation, test endpoints, and assess whether APIs return clear errors. If responses are hard to interpret, they move to another tool.
We built Agent Arena to benchmark how well coding agents complete onboarding tasks across devtools.
Most teams were surprised to see their scores!
Q: Some argue this remains theoretical. Are AI agents really operating at this level now?
AB: In production systems, agents already handle defined tasks.
Developers increasingly run multiple agents in parallel to generate and review pull requests. Anthropic’s Project Vend was an internal experiment where an AI system was put in charge of a small office retail setup handling real transactions.
The question now is whether digital products are designed for agent interaction. If agents become the primary interface to software, companies will need to optimize for machine readability, structured context, and deterministic outputs.
You can also highlight the text and press Ctrl + Enter
