A year ago, Ethan Mollick, the Wharton professor who has become something of a public intellectual on artificial intelligence, made a prediction that rattled the software industry: the era of the traditional software engineer was ending, replaced by something stranger and harder to define. Now, a startup called Wayfound AI is building the infrastructure to make that prediction real — and its CEO, James Cham, has a thesis that’s gaining traction among venture capitalists and enterprise buyers alike.
The core idea is deceptively simple. Instead of writing code line by line, engineers will manage fleets of AI agents that do the building for them. Think less programmer, more project manager — but for software entities that never sleep, never take vacation, and can spin up a prototype in minutes instead of weeks.
Wayfound AI, based in San Francisco, emerged from stealth in early 2025 and has since raised a Series A round that values the company at roughly $200 million, according to people familiar with the matter. As Business Insider reported, Cham’s pitch to enterprise customers centers on a fundamental reframing: your best engineers shouldn’t be writing boilerplate code. They should be orchestrating AI systems that write it for them, intervening only when judgment calls require human expertise.
“The highest-value thing an engineer can do right now is not write code,” Cham told Business Insider. “It’s decide what code should be written, review what the agents produce, and make architectural decisions that require understanding the business.”
That’s a provocative statement in an industry that has spent decades valorizing the 10x engineer — the mythical programmer whose raw coding output dwarfs that of peers. Wayfound’s bet is that the 10x engineer of 2026 won’t be measured by lines of code at all. They’ll be measured by how effectively they direct, correct, and coordinate AI agents working in parallel.
The company’s platform provides what Cham describes as a “management layer” for AI agents. Engineers define objectives, set constraints, and establish quality benchmarks. The agents — powered by large language models fine-tuned on enterprise codebases — then execute. When an agent gets stuck or produces something that fails a test, it flags the issue. The human decides what to do next. It’s a feedback loop, not a replacement.
But make no mistake: the implications for headcount are real.
Several early Wayfound customers, speaking on condition of anonymity because they weren’t authorized to discuss vendor relationships, said they’ve reduced new engineering hires by 30% to 50% while maintaining or increasing output. One fintech company said a team of eight engineers using Wayfound’s platform now produces roughly the same volume of shipped features as a team of twenty did eighteen months ago. The remaining engineers aren’t writing less code — they’re writing almost none. Instead, they spend their days reviewing agent output, designing system architecture, and debugging edge cases that stump the AI.
This tracks with broader industry data. According to a March 2026 survey by Stack Overflow, 61% of professional developers now use AI coding assistants daily, up from 44% a year earlier. But Wayfound is pushing beyond the copilot model — where AI suggests code completions inside an editor — toward full agent autonomy, where the AI handles entire tasks end to end.
The distinction matters enormously.
GitHub Copilot, the most widely adopted AI coding tool, operates as an assistant. It autocompletes functions, suggests snippets, and occasionally writes whole blocks of code. But the human remains in the driver’s seat at all times, typing prompts and accepting or rejecting suggestions keystroke by keystroke. Wayfound’s agents operate differently. They receive a task — “build an API endpoint that handles user authentication with OAuth 2.0 and writes session data to Redis” — and go execute it. They write the code, write the tests, run the tests, fix failures, and submit a pull request for human review.
Cham is careful to distinguish this from the “vibe coding” trend that took off in 2025, where non-technical founders used tools like Cursor and Replit to build quick prototypes by describing what they wanted in plain English. That approach works for MVPs and hackathon projects, he argues, but falls apart at enterprise scale. “Vibe coding is great for getting something that works on a demo,” he told Business Insider. “But when you need something that handles ten thousand concurrent users, passes a SOC 2 audit, and integrates with fifteen legacy systems — you need engineers in the loop. You just need them doing different work.”
The market for this kind of platform is getting crowded fast. Cognition Labs, the company behind the much-hyped Devin AI software engineer, raised $175 million at a $2 billion valuation in late 2024. Factory AI, another competitor, has attracted backing from Sequoia Capital. And the big cloud providers — Amazon, Google, Microsoft — are all building their own agent-based development tools, threatening to commoditize the space before startups can establish defensible positions.
Wayfound’s edge, according to Cham, is its focus on the management problem rather than the generation problem. “Everyone is building better code-generating models,” he said. “We’re building the system that lets you actually run a team of those models in production without everything catching fire.”
That means monitoring, observability, cost controls, and governance. When an AI agent at a Fortune 500 bank writes code that touches customer financial data, someone needs to ensure it complies with regulatory requirements. When an agent racks up $4,000 in API calls to an LLM provider in a single afternoon — something multiple Wayfound customers reported happening during early deployments — someone needs to catch it. The platform includes dashboards that track agent activity, cost per task, error rates, and compliance flags. Think of it as the DevOps toolchain, rebuilt for a world where most of the “devs” are artificial.
The labor market implications are already generating friction. In February 2026, a widely circulated internal memo from a mid-size SaaS company — later confirmed by multiple employees who spoke to The Information — described plans to cut its engineering department by 40% over eighteen months, explicitly citing AI agent adoption as the reason. The memo sparked backlash on social media and renewed debate about whether AI productivity tools are augmenting workers or eliminating them.
Cham’s answer to this is nuanced but ultimately unsentimental. “The number of things that need to be built in software is essentially infinite,” he said. “What’s changing is the ratio of humans to output. Companies that previously couldn’t afford to build custom internal tools will now build them. Companies that had five-year modernization roadmaps will compress them to eighteen months. The demand for software isn’t going down. The demand for a specific type of labor — manual code production — is.”
And the data, so far, supports at least part of that argument. Enterprise software spending hit $1.1 trillion globally in 2025, according to Gartner, and is projected to grow 12% in 2026. Companies aren’t spending less on technology. They’re spending differently — shifting budget from headcount to tooling, from salaries to platform licenses and API costs.
The shift creates new roles, too. Wayfound has coined the term “agent supervisor” for the engineers who manage its AI systems, and several of its customers have begun using the title formally. The job description reads like a hybrid of engineering manager and quality assurance lead: define tasks, review output, debug failures, optimize workflows, ensure compliance. It requires deep technical knowledge but almost no actual coding.
Not everyone is buying it. Sarah Chen, a principal engineer at a major cloud provider who asked that her employer not be named, called the “engineer as manager” framing “a convenient story for selling software.” In her view, the agents still produce too many subtle bugs — the kind that pass automated tests but cause problems in production weeks later. “You still need people who understand the code at a deep level,” she said. “And you don’t develop that understanding by reviewing AI output. You develop it by writing code yourself, hitting walls, and learning from mistakes.”
It’s a fair critique. The question is whether it describes a permanent limitation or a temporary one. LLM capabilities have improved dramatically year over year, and the agents shipping in 2026 are markedly better than those from even six months ago. Anthropic’s Claude, OpenAI’s models, and Google’s Gemini have all shown significant gains on coding benchmarks, and more importantly, on real-world software engineering tasks that require multi-step reasoning and context awareness.
Cham acknowledges the quality gap but argues it’s closing faster than skeptics expect. “Eighteen months ago, you couldn’t trust an agent to write a correct database migration. Now you can, about 90% of the time. In another year, it’ll be 98%. The remaining 2% is where humans add the most value — and that’s exactly where we want engineers focused.”
Wayfound’s business model reflects this bet on enterprise adoption. The company charges per “agent seat” — essentially a license for each AI agent deployed — plus usage-based fees tied to compute and LLM API consumption. Pricing starts at roughly $500 per agent per month for small deployments and drops significantly at scale. For a company replacing twenty engineering hires at an average fully loaded cost of $250,000 each, the math is compelling even at modest productivity assumptions.
The company claims over 40 enterprise customers, including several in financial services, healthcare, and e-commerce. Revenue is growing at what Cham described only as “triple digits year over year,” declining to provide specific figures.
So where does this go?
The optimistic scenario: AI agents become as standard as version control or CI/CD pipelines. Every engineering team has them. Human engineers become architects, supervisors, and strategists. Software gets built faster and cheaper, unlocking a wave of custom applications that were previously too expensive to justify. Wayfound or a company like it becomes the Datadog or PagerDuty of the agent era — the monitoring and management layer that every team needs.
The pessimistic scenario: agent quality plateaus. Subtle bugs accumulate. A high-profile failure — a security breach caused by agent-generated code, a financial system that miscalculates transactions — triggers regulatory backlash and enterprise retreat. The “engineer as manager” model proves to be a degradation of engineering skill, producing a generation of supervisors who can’t actually build anything themselves.
The most likely outcome is somewhere in between, and it’ll vary enormously by industry, company size, and risk tolerance. A startup building a consumer app will adopt agents aggressively. A defense contractor building avionics software will move slowly, if at all.
What’s not in doubt is the direction of travel. The tools are getting better. The economics are getting more compelling. And companies like Wayfound are building the connective tissue between human judgment and machine execution that makes the whole thing work at scale. Whether that’s a net positive for the engineering profession — or the beginning of its hollowing out — is the question that will define the next decade of the software industry.
Cham, for his part, isn’t losing sleep over it. “Every major technological shift has created more opportunity than it destroyed,” he said. “But it’s never been evenly distributed, and it’s never been painless. We’re trying to build the tools that help the engineers who adapt thrive. The ones who insist on doing things the old way — I can’t help them.”
Blunt. Possibly correct. Definitely the kind of statement that makes engineers nervous.
WebProNews is an iEntry Publication