Launch HN: Twill.ai (YC S25) - Get PRs from cloud agents

It runs coding CLIs like Claude Code and Codex in an isolated cloud sandbox for each task. You hand off work via Slack, GitHub, Linear, a web app, or CLI, and it returns a PR, a review, a diagnosis, or a follow-up question, calling you back whenever it needs input.

The limits of running locally are summarized in three points.

• Parallelism: Tasks that touch Docker configs or infra files at the same time don't scale locally, running into conflicts and port reassignment issues.
• Persistence: Closing your laptop stops the agent too, so you can't hand off several tasks overnight and get PRs back in the morning.
• Reliability: Rather than giving an autonomous agent full access to the local filesystem and processes, a per-task sandbox is considered safer.

The initial product was a one-way handoff where you describe a task and get back a PR. It later added a multiplayer feature so a whole team can talk to the same agent, each in their own thread. On top of that, memory was added to preserve rules like "use the existing logger in lib/log.ts, don't use console.log" across future tasks, and automation extended it to event triggers like cron jobs and CI failures.

It's a fiercely competitive space, but rather than building the house themselves, the team chose a strategy of reusing the labs' native CLIs in the cloud. Since the labs keep pouring RL into their own harnesses, the bet is that using those CLIs directly will only get stronger over time.