What did I learn at Code w/Claude London?
And can I string it out to a blogpost?
I was lucky enough to attend Code w/ Claude in London. It didn’t go hugely deep I’m afraid, so I’ve mostly got tid-bits of information that I found useful. So what did I learn?
We’re on an exponential
I think almost every speaker from Anthropic said “we’re on an exponential”. Models are improving at an exponential rate. Unattended task length is improving at an exponential rate. And so on. This was often paired with “AI adoption is linear”. Put the two together and you’ve got an implementation gap, and that’s where we (armed with Claude Code) come into play.
Where does that exponential take us? Well, we can expect better judgement, near infinite context windows and multi-agent co-ordination. We’ve gone from completion to agentic execution. The next step will be continuous, proactive AI with always-on agents. Prompts will move from descriptions of what to do, to the outcomes we need to achieve (as general as “keep the project on track”.
Designing apps
Three guiding policies.
Build for the next model - as models get smarter, previous generations of scaffolding often hold it back. The direction of travel is more general tools, rather than highly specialize ones.
Curate frontier evals - You need to understand how your system performs with models (this was a common theme(
Treat model upgrades as a feature
What model should you use?
You should use the model that performs best on your eval. You don’t have evals? Use the defaults for the best results.
The /advisor mode was recommended as a way to achieve near frontier performance at five times less cost. It uses the Sonnet model by default and escalates to Opus when needed.
Some Claude Code Tips and Tricks
/remote-control - Control Claude Code sessions from your phone so you don’t have to walk around with your laptop open.
/tui fullscreen - Flicker-free rendering (rendering is virtualized, no more horrible flicker)
claude agents - Manage a team of agents. Anything you type spawns off an agent and you can jump in/out of them anytime. Very nice!
/routine - Run remote agents to respond to events (or just a cron job). For example, sweep through the tech debt backlog, update code-reviews and so on.
/ultrareview - Launch a multi-agent review of your pull request. Used by Anthropic on every PR.
claude --worktree - Claude can use git worktrees to allow multiple agents
Some new things
Alas, there was no new model for us to play with.
Self-hosted sandboxes do what they say on the tin! I can see this being super useful for bringing agentic workflows in house when you’re worried about your data leaving your infrastructure.
MCP tunnels solve the problem of securely connecting Claude to MCP servers running in your network (without opening your MCP server to the internet!).
How do you run an AI-native org?
By far my favourite talk because some of the opinions were the opposite of what I thought.
The bottlenecks have moved. The old bottlenecks were code, tests and refactoring. AI can do that now. The new bottlenecks are verification, review, security and XFN (I’m using the acronym to appear hip, I believe it’s x-functional).
So far, so standard. But what I found interesting was how the norms are changing at Anthropic.
There’s less planning. You might think that the emphasis should be there (after all, giving AI a good spec feels like it should be a good idea),
Onboarding is easier. The cost of answering questions is zero; AI can answer questions about code.
Technical debates become show don’t tell. No more arguing about design A vs. design B when you can just implement. In Technical Debates, code wins.
Less design doc rituals. No theatre of writing down X, just build it instead.
The source of truth is the codebase. All documentation in the codebase.
Code review is the problem we all thought it would be. Anthropic use Claude for style, obvious bugs, spec drift and so on, but humans where there’s legal concerns, high-risk, trust-boundaries or just product taste.
What does the team make up of the future look like? A combination of creative builders with product sense, and others with deep system expertise for the hard parts. All managers at Anthropic are ICs first, contributing code. And everyone dog-foods whatever you are building.
How’d you roll out this?
Claude-ify all your workflows - this frees up humans for the harder stuff
Tell people they can delete processes
You’ll be winning if onboarding is quicker, PR cycle time is faster and the number of Claude-assisted commits is going up.