The bottleneck isn't code anymore

I’ve received 10s of DMs about how we’ve changed the way we work over the last week so I figured I'd write the longer version.

At Mio, we're a 5-person team building a context tool. We’re pre-PMF, experimentation velocity is our north star. At the start of the month, we went agentic: changing how we work by making agents, not engineers, our primary contributors.

Here's what actually happened.

What's working

4.2x more PRs closed this week vs. late January. 117 vs. 28 for the same team size.

We don't write or review code. Cursor is out, ADEs are in. We define intent, manage agents, and evaluate output.

We prototype in code. No wireframes. No ideas docs. Our design system lives in GitHub (design tokens). If we're debating between two approaches, we just build both and pick the better one. A/B testing used to be a luxury for early stage startups, now it's faster than arguing about it and creativity went up.

What's changed

Over 90% of tickets created are resolved within 2h. Prioritisation used to happen across a backlog in linear. Now it happens in real-time, we log decisions and move on.

No roadmap and multiple adjustments a day. We scrapped the 3 month roadmap, we navigate on a weekly basis. On mondays, we align on the outline of a project and adjust it live as we collect insights.

Raise the quality bar. When code is free, "good enough" stops being acceptable. We moved from MVP to Minimum Lovable Product. If it doesn't delight, rebuild it. Faster iterations require higher standards.

What's harder

The PM is the bottleneck. When agents can build anything in hours, the limiting factor is knowing what to build: collecting representative data and making sense of it. Engineers are now PMs.

Scope creep is a new kind of problem. When anything can be built extremely fast, the temptation is to build everything. Staying focused requires discipline: the skill isn't building, it's deciding what not to build.

Coordination cost increased. Code review wasn’t just about code quality, it was also a coordination mechanism. Now that no one reviews code, new systems are needed to ensure a shared understanding of the codebase.

Context bloat. It takes more time for humans to read than it takes for models to write. No one wants to spend their time reading AI generated docs. We put in place a subtle but important rule: all docs start with a human-written TLDR containing the insights and decisions. The rest of the doc is produced by agents to be consumed by agents.

The second order effects

Here's what became obvious after a few weeks: the bottleneck isn't code, it's knowing what to build.

When everything ships in hours, the quality of your decisions becomes the only thing that matters. And decisions are only as good as the context behind them. That's where it breaks down.

Information is scattered across Slack, Linear, Notion, GitHub, email, meetings. No single person can synthesize all of it fast enough. You end up with fast execution on top of slow understanding. That's worse than being slow everywhere, because you ship the wrong thing with confidence.

If this resonates, DM me. We have a shared context product you should try.