A Year of Claude Code: What Actually Changed in How I Build
Claude Code launched a year ago to a muted reaction. Now people direct fleets of agents, sometimes from their phones. The two big leaps, the habits that stuck, and what it means for how we work.
Guillaume Rufenacht
AI Product Manager · Lisbon
Claude Code launched a little over a year ago to a muted reaction, a short video, a couple of excited replies. A year later, the people building with it don’t write code so much as direct fleets of agents, sometimes from their phones. That’s two enormous leaps in twelve months, and the trajectory says more about where building is going than any single feature. Here’s what actually changed, and what it means.
I’ve been building with it the whole way, from the AI SDR system to the pipelines behind Geonimo, so this is the arc as I’ve lived it.
Key takeaways
- The unit of work went from one agent you supervise to dozens or hundreds you direct.
- The reliable pattern: fix the system (a rule or skill) when the agent errs, so it can run unattended.
- Trust earned through good defaults (auto-accept with a safety classifier) is what made fleets possible.
- Routines, agents that proactively pick up issues and fix them, turned the agent SDK from idea into daily use.
- Roles are merging: PMs, designers, and analysts all build now, and judgment beats typing speed.
Two leaps in a year
The first leap was “I don’t write the source code, I talk to an agent that writes it.” The second, happening now, is “I don’t talk to one agent, I direct many, or set up a routine that prompts them for me.” A year ago, working with Claude Code meant six terminal tabs and one task at a time. Now it’s an agent view managing worktrees, several agents running at once, and, remarkably, a real share of the work kicked off from a phone via remote control. The form factor keeps changing because the models keep getting more capable, and capability buys autonomy.
The habits that stuck
Through all that change, a few principles held, and they’re the ones worth internalizing because they’ll outlast any specific UI:
Fix the system, not the instance. The single most important habit: when the agent makes a mistake, encode the fix in a durable place (a rule, a skill) so it never recurs. That’s what lets an agent run for a long time without you.
Trust through good defaults. Moving from approving every action to auto-accept with a model checking safety was the unlock for running many agents at once, and it’s counterintuitively safer than rubber-stamping prompts your eyes have glazed over. I unpack both habits in how Anthropic’s team uses Claude Code.
The routine is the new agent
The idea that turned “you can use Claude Code programmatically” into something people actually rely on is routines: agents that proactively watch for work, an incoming bug report, a failing check, and put up a fix before a human even looks. The uncanny moment is when your own agent tells you someone else’s already fixed the thing you were about to. Code review, CI babysitting, rebasing, the chores that used to fill a day, fade into the background. The leap is from talking to an agent to delegating to a loop.
The bigger shift
What it means for how we work
When the agent writes the code, the scarce skill stops being typing and becomes judgment: product taste, clear problem framing, knowing what’s worth building. That’s why the roles are merging, designers, PMs, even finance and data folks now build directly, and engineers ship end to end. It rewards exactly the profile I’ve leaned into: someone with product judgment who can also build. A year in, the most striking thing isn’t how much the tool can do, it’s how much the job has changed around it.
The takeaway
I build this way every day, see building with Claude Code, what I’ve shipped, or get in touch.
Frequently asked questions
How has building with Claude Code changed in a year?
From writing source code, to talking to one agent that writes it, to directing many agents or setting up routines that prompt them, two large leaps in twelve months, with a real share of work now started from a phone.
What habit matters most for working with Claude Code?
Fix the system, not the instance. When the agent errs, encode the correction in a durable rule or skill so it never recurs. That's what lets an agent run unattended.
What is a routine in this context?
An agent that proactively watches for work, an incoming bug report, a failing check, and puts up a fix before a human looks. Routines turned programmatic agent use from an idea into something people rely on daily.
Does AI mean engineers aren't needed?
The opposite of the scare story: when the agent writes code, the scarce skill becomes judgment, product taste, problem framing, knowing what's worth building. Roles are merging, and that profile is more valuable, not less.
Work with me
Want a system like this built for your pipeline?
I help teams take AI from a clever prototype to dependable production, outbound engines, lead intelligence, and the LLM pipelines underneath. See what I have shipped or get in touch.