Introduction

​

Software is being built by agents

The way software gets built is changing fast. Coding agents are no longer autocomplete tools that suggest the next line — they’re autonomous systems that run for hours, write thousands of lines of code, and ship production-ready pull requests. Cursor’s long-running agents routinely run for 25–36 hours on tasks like full architecture overhauls and mobile app implementations, producing PRs with merge rates comparable to human-written code. Claude Code’s agent teams coordinate multiple independent agents working in parallel — investigating competing hypotheses, building separate modules, and synthesizing results across workstreams. Engineering teams are already seeing this play out at scale:

• Ramp reports that 57% of merged PRs in a recent 24-hour period came from their background agent, up from 30% just weeks earlier.
• Stripe built “Minions,” one-shot end-to-end coding agents responsible for over a thousand PRs merged each week.
• Coinbase hit 5% of all merged PRs coming from in-house background agents, with adoption growing exponentially.

These teams spent months building custom internal infrastructure. Magnet gives every team the same capabilities, out of the box.

​

What is Magnet?

Magnet is the multiplayer workspace for building software with autonomous, long-running coding agents. It gives your team a shared environment where agents run in isolated cloud sandboxes — with full repository access, real-time streaming, and automatic pull request creation. The same way you give your engineers the best tools and context to do their work, Magnet lets you build the best agents for your team — personalized to your codebase, your workflows, and your standards.

​

Key capabilities

​

Long-running agents need a workspace

Agents that run for hours or days can’t live inside an IDE tab. They need their own environments, their own compute, and a shared place where your team can see what’s happening. Magnet provides all of this. Each session runs in a full cloud sandbox with your repository cloned, dependencies installed, and your dev server running. The agent can read and write files, run shell commands, execute tests, and make git commits — just like a developer would. You can kick off a dozen sessions in parallel without tying up your laptop, check back hours later, and review the results. As agents grow more capable — running longer, coordinating in teams, tackling bigger projects — the infrastructure around them matters more, not less. Magnet is built for this future: a workspace where you manage intent and quality while agents handle implementation.

​

Programming is becoming specifying

The shift is already underway. Programming is becoming less about writing code line by line and more about specifying what you want, clarifying intent, and reviewing results. The best teams will be the ones who learn to direct agents effectively — writing clear prompts, structuring good issues, and building environments that set agents up to succeed. This changes what a software workspace needs to be. It’s no longer enough to have a code editor and a task board in separate tools. The workspace for managing a software project will be far more integrated than anything that exists today — issues, specs, docs, diagrams, coding agents, deep integrations to your codebase, and autonomous learning systems, all in one place. We’re building Magnet to be that workspace. We’re exploring the new workflows and architecture for intent and quality — helping teams evolve their software systems with agents as first-class collaborators, not just tools you invoke from a terminal.

​

Build your competitive advantage

Every codebase is different. Generic tools treat every project the same — but the teams seeing real results are the ones building agents that understand their code. With Magnet, you configure environments with your specific setup scripts, dependencies, and dev servers. Your agents run with the same context your engineers have: your repo structure, your test suite, your coding standards. The more you invest in your agent setup, the better the results — and the wider the gap between your team and everyone else. This is the path to self-building software. Not a single leap, but an iterative process: set up an agent, let it tackle a task, review the results, refine the setup, repeat. Each cycle, the agent gets more effective. Each cycle, your team ships faster. Magnet gives you the platform to run that loop continuously.

​

How it works

1. Create a session — provide a prompt and optionally link a repository.
2. Agent executes — Claude Code runs in a sandboxed environment, cloning your repo and making changes.
3. Real-time streaming — watch the agent think, write code, and use tools as it works.
4. Review & merge — the agent opens a pull request. Review the diff in Magnet or on GitHub.

​

Use cases

Long-running coding agents unlock workflows that weren’t possible before:

• Bug fixes on the spot — see an issue in production? Describe the fix and let the agent ship a PR while you move on.
• Parallel exploration — kick off multiple sessions with different approaches and see which one lands.
• Large features and refactors — hand off multi-hour tasks and come back to a working PR, not a half-finished branch.
• Delegating scoped tasks — validation, tests, refactoring a module — define the task and let the agent handle it.
• PR review follow-ups — instead of leaving a comment and waiting, describe the change and get a PR in minutes.
• Overnight work — start a session before you leave for the day, review the PR in the morning.

​

Next steps