What the summit was actually about.

Code with Claude 2026 in London on Tuesday 19th May was framed as a coding event. It wasn't. Once you strip out the demos and product launches, every serious session was about the same thing: what breaks inside a company when models stop being the constraint. When the model is the slow part of a system, everything else hides behind it. When the model speeds up, every other part gets exposed as the new floor. The platform talks pointed at it. The enterprise panel said it out loud. This deck is my reading of where the constraint went.

The task horizon keeps doubling.

A year ago Claude could reliably work for minutes. Today most users have agents running for hours. Jeremy Hadfield's "The capability curve" on the main stage put the doubling on screen. The Anthropic team is openly building toward continuous, always-on agents owning goals, not tasks. Lisa Crofoot's line for the developers in the room was the only line that mattered: build for the next model, not the current one. Most companies are still doing the opposite.

For decades, coding was the constraint.

The whole product playbook assumed engineers were the scarce input. Roadmaps were carved to fit capacity. Ideas got binned because nobody could ship them. That bottleneck is loosening. Niklas Gustavsson, Spotify's VP and Chief Architect, put it plainly on stage: coding is much less of a bottleneck now, and the constraints are moving somewhere else. The interesting question is no longer how fast can we ship. It is which constraint replaces it.

Spotify's PR frequency, after Opus 4.5.

The bottleneck isn't holding. It's gone.

Humans pick what to read.

PR review was built for a world where humans wrote every line. Agents now write most lines, and review queue depth becomes the floor on shipping. The answer is not more reviewers. It is a shift in what humans actually read: auto-merge on low-risk diffs, human eyes for the judgement calls. Spotify and Anthropic are already running this model in production. Spotify said it on stage: PR frequency is up 76% since Opus 4.5 went into the codebase.

Now we have to decide what to ship.

Spotify said it on stage: when coding stops being scarce, the constraint moves to deciding what we ship to users. Alex Kaluzny, Doctolib's CTO, described how prototyping in their actual production codebase used to take weeks. It now takes minutes. The result is not faster product. The result is too many viable ideas, and a leadership team that cannot tell them apart fast enough. Alexander Bricken's "The thinking lever" research talk pointed at the cognitive side of the same problem. Taste, conviction, and a willingness to kill ideas become the new floor on output.

Who is allowed to do what.

Identity is the unsolved problem of agent-native systems. An agent spawns another agent, calls a service, writes a record. Who has permission. What is audited. What does the chain look like. Ruslan Semenov said on the enterprise panel that monday.com is rebuilding its permission model to treat agents as first-class users. Anthropic answers from the other side: Claude Managed Agents gives each agent a scoped identity. There is a protocol war coming, and the orgs that wake up to it will own the standard.

A consistent codebase is now an AI feature.

Spotify spent a decade pushing engineers toward a smaller set of approved technologies, a single developer portal in Backstage, and lint rules that tell the model when it has drifted from the house pattern. What started as a developer experience play turned out to be the strongest moat for AI agents in their codebase. Fragmented code performs worse with Claude. Consistent code performs better. The implication runs deeper than developer experience.

Last year's scaffolding is this year's drag.

The most under-reported line came from Anthropic's research team. Sid Bidasaria's "Stop babysitting your agents" and Margot van Laar's "The prompting playbook" picked at the same scab from different angles. The harnesses, loops and tools you built around earlier models still sit in your stack. As Claude gets smarter, that scaffolding holds it back. Generalised primitives beat bespoke wrappers. I have lived this on my own stack: 28 hooks across my workflows, several rewritten or retired as new models needed less of them.

The platform isn't about models. It's about infra.

Each platform release this year answers a bottleneck. Claude Managed Agents: scoped agent identity (bottleneck 3). Ravi Trivedi's "Memory and dreaming for self-learning agents": agents that retain context (bottlenecks 1 and 2). Ralph Ramos's "What's new in Claude Code" and Punit Shah's "Getting more out of the Claude Platform": a codebase the model can reason about (bottleneck 4). The serious money in 2026 is in operating agents at the org level, not in calling them.

"But what if I'm not Spotify?"

Fair. None of the named orgs in this deck are small. The honest version of the answer is: the bottlenecks scale down. Review becomes who validates one agent's output, not who drowns under a thousand PRs. Identity becomes what your single agent is allowed to do, not an enterprise auth chain. Standardisation becomes how consistent your own code is to the model that reads it. The names change. The constraint moving doesn't. If anything, small teams feel it sooner, because there is no platform team to absorb the shift for you.

Three falsifiable checks.

Every reading puts the author at risk. Three checks would update this read.

Q4 2026: are PR-review surveys still ranking bugs above PR volume?
Mid-2027: are product teams still evaluating throughput over judgement?
May 2027: did per-token compute pricing reverse?

If any resolves the wrong way, the thesis bends or breaks.