The first wave of AI coding tools lived inside a chat box. You asked, it answered. You closed the tab and the agent forgot you existed. That model is fine for snippet generation, but it’s a bad fit for actual software work, where context — who’s on the team, what tickets are open, why this PR was rejected last month — is most of the job.
What I want to talk about is what’s quietly replacing that pattern: persistent coding agents that act less like assistants and more like coworkers. Two different tools convinced me this is real: Claude Code on the IDE side, and Paperclip on the org side.
Claude Code: the agent that lives in your repo
Claude Code is Anthropic’s CLI agent, and the thing that makes it different from a chat-window LLM is that it runs inside the project. It reads files with the same paths you do, runs the same npm test, hits the same dev server on localhost:5173. When it edits, the diff lands on your branch. When it commits, your hooks run.
That sounds mundane until you’ve used it for a week. Suddenly the agent isn’t a thing you copy snippets out of — it’s a thing you assign tasks to, the same way you’d assign them to a junior engineer who happens to type 200 wpm and never gets bored grepping. The mental shift is bigger than the technology shift. You stop asking “can the model do this?” and start asking “is this work I want to delegate?”
The thing Claude Code gets right is that the IDE is the interface, not the product. The product is the agent’s ability to act inside your existing tools — git, your test runner, your editor’s file tree, your terminal. No portal, no separate dashboard, no copy-paste loop. Just the loop you already have, with a teammate in it.
Paperclip: the agent that lives in your company
Paperclip takes the opposite slice. Instead of dropping an agent into one repo, it wraps agents into roles. You spin up a “company” with a CEO, a CTO, a PM, an engineer, a QA — each one is a Claude-powered agent with its own instructions, its own ticket queue, its own memory. Tickets get assigned, comments get left, work gets handed off. There are heartbeats: the agent wakes up on a schedule, pulls its assigned issues, and works them.
The interesting bet is that the interesting unit of automation isn’t the function call — it’s the org chart. A single agent doing a single task is just a smarter autocomplete. A graph of agents that escalate to each other, ask each other for review, and collectively own a backlog is starting to look like a small team. And teams compose differently than tools do: they have politics, blockers, reviews, deadlines. Modeling those explicitly turns out to be more useful than modeling them implicitly inside one giant prompt.
What I like about Paperclip is the discipline it forces. Because every action runs as a specific role, you can’t wave away accountability. The CTO agent is the one who shipped the bad migration. The QA agent is the one who let it through. You can read the ticket trail. You can change the role’s instructions and replay.
Why both matter
The combination is the point. Claude Code is the hands: an agent that can actually move bytes inside a repo. Paperclip is the org: a structure that decides which hands get which work, and what happens when work is done. Bolt them together and you stop having “an AI feature” and start having something that looks like staffing.
There’s a real question hiding here about where humans fit. My current answer: the human is the founder, the architect, and the customer. The agents are the staff. You set direction, you read the ticket queue at the end of the day, and you redirect when the org drifts. That’s a meaningfully different job than “prompt engineer.” It’s closer to “manager who can also code.”
Persistent agents are not coming. They’re here, and they’re already cheaper than they should be. The tools that win the next two years will be the ones that treat agents as members of an organization, not as features inside a product. Claude Code and Paperclip are early, opinionated answers to that — and worth the time to actually use, not just read about.
