As part of The State of the AI Frontier 2026, AI Frontier Network invited leaders building and deploying AI in the real world to share where the frontier is actually moving. In this contribution, Andrii Yasynyshyn, CEO & Co-Founder at Ralabs, gives a candid read on what changes in 2026 — and what to watch.
On the shift that defines 2026
The shift is from improvising with agents to trying to run a real method. For two years everyone hand-rolled their own way of working with coding agents. Now the contenders for a shared method have arrived: GitHub's Spec Kit, AWS's Kiro, Anthropic's Claude Code skills, plus Tessl, OpenSpec, and BMAD. They converge on one idea, spec-driven development: an engineer writes a precise spec, and agents build from it. Two things keep this honest. First, it changes how engineers work, it doesn't remove them. Someone still writes the spec, owns the architecture, and decides whether the output is right. Second, it isn't proven, and it isn't automatically cheaper. Only about 13% of developers use AI across the whole lifecycle today, and only a third trust what it produces. Used carelessly, agentic development can cost more than a capable human team, because an inefficient workflow burns tokens fast. So my real answer is a test, not a prediction: 2026 is the year we find out whether spec-driven development becomes a genuine paradigm or turns into another well-funded experiment that doesn't stick. By December we should know which.
On the real unlock
Ads inside AI answers are already here and already priced. ChatGPT's ads hit roughly $100M annualized within weeks of launch, and Google shows ads in about a quarter of its AI Mode results. That part is obvious. What the market hasn't priced is the layer underneath: paid influence over what an autonomous agent decides, not what a human reads. As agents start doing the buying through protocols like OpenAI's ACP and Google's UCP, the "ad" stops being a labeled box and becomes a thumb on the scale of a recommendation you never see. Today you can't ask why an agent chose one vendor over another, or whether that choice was bought. The countermeasure that doesn't exist yet is an integrity layer for agents, the equivalent of an ad-blocker, that detects and discloses sponsored influence and resists prompt-injection-style manipulation. Whoever builds the trust layer for agent commerce is sitting on something the market is ignoring.
On the trap to avoid
Faster typing isn't faster business.
The second trap is assuming AI replaces engineers and cleans up your existing systems. AI is strong on greenfield, where it starts fresh, and weak on the systems most companies actually run: long-lived, standard codebases that have grown and been added to for years. Working in those takes context and judgment that AI doesn't have on its own. The third trap is price, and it's the one people miss. AI doesn't automatically come out cheaper than a developer team. Two reasons. First, talking to people is efficient: a developer understands what you mean from a few words and fills in the rest. An AI agent needs every detail spelled out, and each detail you leave out is paid for twice, once in tokens, and again in your time when the result misses and you redo it. Second, frequent, fast-changing work makes that worse, because each change means re-specifying everything for the agent. So replacing people with AI can land you in a reality where you talk to the agent more than you would to a person, and pay more for the privilege. And for all the noise, the models barely moved this year. Coding with Opus 4.8 today feels about the same as 4.6 in February. Capability isn't the constraint. Knowing what to build, and what it truly costs to build it, is.
On the organization that adapts
AI is changing two things for us: the roles we add, and how our existing engineers work. The new roles are real: AI engineers, and AI workflow engineers who design how agents and people hand work back and forth. But the bigger change is the mindset shift for the developers we already have. The old split was roughly 80% writing code, 20% thinking about the product. That's inverting. The work is becoming 80% understanding the product, the requirements, and the constraints, and 20% reviewing what the agents produce against them. We're moving people from programmers to product-minded engineers. The same engineers who were strong before are strong here, they're applying the same judgment higher up the stack. What humans own is narrowing to the parts that carry consequence: deciding what to build, owning the outcome, and catching work that isn't right before it ships. Everything that can be specified and delegated to an agent, is. What can't, stays human.
On the benchmark that matters
The signal I'd watch: the share of companies that can show a measurable profit-and-loss impact from AI. Not usage, not "productivity," actual money. MIT put it at 5% in 2025. Spending is set to jump about 65% this year, from roughly $7M to $11.6M per enterprise on average. My call: by December 2026 the share showing real P&L impact is still in single digits, under 10%, even as spend climbs. The gap between AI investment and AI return widens before it closes. Cash is the only scoreboard that matters here. If you're building something brilliant that doesn't move revenue or margin, you're funding a demo instead of a business.
