AI-Assisted Coding Fatigue Sparks Debate at Berlin Conf

The morning after: thirty PRs, one maintainer

At Local-First Conf in Berlin this week, developers traded stories that sounded less like triumph and more like triage. Douwe, who maintains the Pydantic AI framework, described waking up to thirty pull requests every morning, each pulled overnight by someone's AI agent. His question to the room wasn't about productivity gains — it was existential: if AI review gets delegated further, "what am I still doing here?"

That sentiment threaded through much of the conference. Douwe also described a loss of collaborative learning, saying contributions now disappear into "some AI black hole," with no person actually learning from the exchange. Marcelo, another attendee, joked about running five Claude Code sessions simultaneously — partly so he wouldn't notice when one froze.

Projects are starting to say no

Not every open-source community is absorbing the volume quietly. According to the report, projects like Zig and Gentoo are refusing to accept LLM-generated PRs outright. Maintainers elsewhere describe being overwhelmed by the sheer number of AI-assisted submissions, raising real sustainability questions for review workflows that were never designed for this scale.

The human cost shows up outside the repo, too. A Pydantic article referenced at the conference described the author staying up until nearly 2am recently, prompting an AI to get a plan right. The same article argues that LLM-assisted programming has replaced the dopamine of creative work with the cognitive load of review and supervision — and that coding with an LLM is "an intensely solitary activity" that can chill communication between developers.

Builders responding with new tools

Not everyone at the conference was pessimistic. Armin Ronacher — creator of Flask and an early Sentry team member — has founded a company called Earendil, which builds Pi.dev, an open-source coding agent harness. Its stated purpose: "In a world hurtling towards AI, we believe humans are the best agents." Ronacher said Earendil's harness auto-closes almost all PRs and issues automatically — though the report doesn't specify exact figures or how the company is funded or staffed.

The Pydantic article itself struck a more optimistic note, stating that the gap between AI tool promise and reality has "narrowed to something meaningful" — suggesting the frustration may coexist with genuine progress.

A geopolitical wrinkle

Adding to the uncertainty, the US government cut off non-US citizens' access to Anthropic's latest frontier model a few weeks before the conference. On stage, Martin Kleppmann noted that the probability of a US-Europe conflict, while still low, was effectively zero just a year ago — a reminder that access to frontier AI tools may not be geopolitically neutral going forward.

Why founders should care

Maintainer overload from AI-generated PRs suggests there may be growing demand for tooling that triages or auto-manages contributions at scale — Earendil's approach is one early example.
Policies at Zig and Gentoo rejecting LLM PRs hint that founders building AI-assisted dev tools should probably think carefully about contribution norms and community trust, since backlash isn't uniform but isn't rare either.
The described burnout and loss of collaborative learning could signal an opening for products that reduce review burden without fully removing human oversight — a middle ground between full automation and status quo review.
Restricted access to Anthropic's frontier model for non-US citizens is a concrete signal that geopolitical risk is now a real variable for startups building on a single AI provider, particularly for international teams.
The narrowing gap between AI promise and reality, as described in the Pydantic article, suggests the underlying capability is maturing — even as the human and organizational costs of adopting it remain unresolved.

What's still unclear