AI-Assisted Development Looks Easy, That's the Problem - ft. Daniel Eneström and Daniel Astudillo
In this episode of the Accedo Playback Podcast, we had a chat with Daniel Eneström, Head of Engineering for Europe and Latin America, and Daniel Astudillo, Head of Engineering for North America, to discuss the impact of AI on the streaming software industry, focusing on engineering challenges, expertise, and the evolving development landscape. AI has significantly lowered the barrier to building streaming software, allowing the first 80% of a project to be completed quickly. However, this comes with real costs: more security vulnerabilities, more code duplication, and engineers who understand less of what they've actually shipped. The conversation turns on what closes the remaining 20%: judgment built from hundreds of deployments, the kind that knows how systems fail together during a live event or why a phone experience needs different decisions than a living room one. Our guests also dive deep into the actual practices: managing shared context across teams and agents, what makes an effective agent harness, and why a hybrid of frontier and open-source models is likely where the industry lands as token costs rise. Topic discussed 01:10 What's actually changed in AI-assisted development over the last two years 02:28 Every team reaches the same 80% — and why that's no longer the differentiator 05:32 The hidden cost of AI-assisted code: longer PRs, slower reviews, lower comprehension 09:36 What "accumulated expertise earned over hundreds of deployments" means in practice 13:04 Managing shared context across teams and AI agents at scale 18:04 Building an effective agent harness: why inputs are king 22:05 Frontier models vs. open-source tooling, and the case for a hybrid approach 26:00 Is genuine differentiation still possible once AI raises everyone's floor? 29:20 The one question technical leaders should be asking but usually aren't