Copilot or Autopilot? The AI Architecture Decision That Defines the Next Decade

Inbal Shani, then chief product officer of GitHub, defends the copilot model with a systems-thinking argument rooted in more data about developer behavior than anyone else has. As CPO overseeing GitHub Copilot -- with 1.5 million developers and over 37,000 organizations -- she has observed at scale how humans actually use AI assistance.

"Copilot is a copilot. It's not a pilot. You still need the human in the loop. But that means that now the software developer or the user of the AI tools to develop software needs to form a different thinking."

Inbal Shani

▶ 00:08:02

Shani's argument is that the hardest part of software development is not writing code but understanding systems, architecture, and connected experiences. Those are precisely the capabilities AI cannot yet replicate. Copilots preserve human judgment where it matters most while automating the mechanical work.

More importantly, she argues that copilots elevate rather than diminish the human role. When AI handles code generation, junior developers can "spend more time from the get-go understanding the system, understanding the environment that they're building, understanding the product that they're building, which today they don't have time because they are still learning how to code." The copilot model does not just make developers faster -- it makes them think more like senior engineers from day one.

"Let me be very clear, you cannot cut your people. You have to have a human in the loop. Copilot is a copilot, it is not a pilot. And I keep on saying that sentence again and again."

Inbal Shani

▶ 00:12:23

Her data supports the claim: developers write code 55% faster with Copilot, 85% feel more confident in code quality, and crucially, the most elite engineers love it just as much as juniors do. It is not a crutch -- it is a genuine multiplier at every skill level.

The temporal caveat matters here: Shani's interview is from December 2023, before the explosive growth of agentic coding tools throughout 2024-2025. GitHub itself has since moved significantly toward more agentic features, and competitors like Cursor ($300M ARR in two years) have proven that developers want more autonomy in their tools, not less. The copilot framing may have been correct at the time while simultaneously being a transitional stage.

Scott Wu, co-founder and CEO of Cognition, represents the most aggressive position. Wu built Devin, the world's first autonomous AI software engineer, and his 15-person team runs one of the most agent-native engineering organizations in existence, with each engineer managing up to five Devins simultaneously.

"Devin is a fully autonomous software engineer that is going to work on tasks end to end... you can tag Devin on an issue in Slack, you can tag Devin in Linear, and Devin will make pull requests in your GitHub, and so it's very much built to work with engineering teams as your junior engineer."

Scott Wu

▶ 00:06:10

The critical difference from copilots is the interaction model. You do not pair with Devin in real time. You delegate to it asynchronously, the way you would assign a task to a human engineer. The best workflow, Wu found, is "to work with multiple Devins and to run them asynchronously and to kick them off and to only jump in basically as you needed to provide feedback or steer the plan." This is delegation and review, not collaboration.

Wu is honest about the limits. He embraces the concept of "jagged intelligence" -- agents are dramatically better than humans at some tasks and dramatically worse at others. Devin evolved from "a high school CS student" at launch to "a college intern" to "a junior engineer" over the course of a year. The trajectory matters more than the current state.

"People didn't really believe that an agent was possible. Right. And it was, I mean, it was a very different time."

Scott Wu

▶ 00:07:02

Cognition's own metrics tell the story: about a quarter of their production PRs are authored by Devin today, and they expect this to exceed 50% by year end. But Wu identified a crucial pattern in customer success: companies that thrive with Devin follow the same onboarding process they would use for a new human engineer. Start with small tasks, let the agent get familiar with the codebase, set up the environment, then gradually increase scope. Customers who failed gave Devin massive re-architecture tasks on day one -- just as a new hire would fail at those.

Eoghan McCabe, founder and CEO of Intercom, provides the most compelling existence proof. McCabe founded Intercom 14 years ago, grew it to a multi-billion-dollar valuation, then watched it approach negative growth before transforming it into the market-leading AI agent company. Fin, their customer support agent, is growing at 300%+ year over year, leading the market by every metric.

"Fin is our AI agent who will pass 100 million ARR with Fin in less than three quarters."

Eoghan McCabe

▶ 00:00:15

McCabe's most revealing insight is about pricing. Intercom historically had one of the most hated pricing models in SaaS -- complex, opaque, and widely mocked in viral memes. For Fin, they made a deliberate philosophical choice: charge $0.99 per resolution so that revenue was "100% aligned with the value that they attained." Early on, each resolution cost Intercom $1.20 to serve, meaning they lost money on every transaction. McCabe bet that inference costs would drop. They did.

"We charge 99 cents to resolve tickets, customer problems, and we have a higher resolution rate than anyone else... It is the metric by which these agents are assessed, and we wanted our revenue to be 100% aligned with the value that they attained."

Eoghan McCabe

▶ 00:23:33

The $0.99 price was strategically chosen. Their research showed businesses typically spent $20-30 per human-resolved ticket. But McCabe sensed customers would "not value the digital work as much as the human work, even though the digital work is better." Sub-dollar pricing removed the objection entirely: if you will not pay 99 cents for an instant, expert resolution available 24/7, there is no business here.

McCabe also challenged his own history. Intercom's original mission was "make internet business personal." Now they sell an AI agent. His resolution: "Providing a customer with a highly engaged, instantly available expert, consistent, fast, charismatic, funny, friendly, personal agent available for literally every single customer every minute of the day around the clock is so much more personal than making them wait 2, 3, 4 days for a crappy canned response."

Aparna Chennapragada, chief product officer at Microsoft, offers the most analytical framework. Previously CPO at Robinhood and VP at Google (Google Lens, Search, AI Assistant), she brings experience across consumer and enterprise AI at massive scale.

Rather than picking a side, she defines agents along three dimensions: autonomy (how much you can delegate), complexity (multi-step vs. one-shot tasks), and natural interaction (conversational rather than GUI-based). This reframes the debate from a binary to a continuous spectrum.

"When I think about agents, I think about three things. One, it's autonomy... And it's a spectrum, it's not a zero-one. Second, I think of as complexity. It's not a one-shot, 'Hey, summarize this document,' but it's 'build me this prototype.' And then the third one I think of is it's a much more natural interaction."

Aparna Chennapragada

▶ 00:17:10

Her position on pricing is equally nuanced. She identifies three coexisting models -- seat-based, usage-based, and outcome-based -- and argues at least two should coexist.

"We've just barely scratched the surface of whether you do seat monetization, usage like on tap, and then of course outcome-based stuff... So all three to me are kind of like, great, but at least two out of three should coexist."

Aparna Chennapragada

▶ 00:43:24

She also introduces a concept no one else raises: NLX -- Natural Language Experience as the new UX. Agent products are not just "chat with AI." Conversations have grammars, structures, and invisible UI elements. Plans, progress indicators, follow-ups, and editable workflows all need to be designed as deliberately as any GUI. The companies that master NLX design will build durable advantages regardless of where their products sit on the copilot-to-agent spectrum.