About
We built an engineer, not an autocomplete
Elisha started as an internal tool for a team with one engineer and a queue that never got shorter. It turned out the interesting problem wasn't writing the code — it was everything either side of writing the code.
The gap we kept falling into
Coding assistants are good at the part a developer was already going to do. They make the person in the editor faster. But on a small team, the person in the editor isn't the bottleneck — the queue is. The bug reported in #support sits for two days because the only person who can fix it is halfway through something bigger.
An assistant that returns a snippet doesn't shorten that queue. Someone still has to open the repo, find the right file, apply the change, run the suite, and push it. We kept watching that leftover work swallow the time the assistant was supposed to save.
So we moved the finish line
A pull request is the smallest honest unit of work. Anything less leaves the job on your desk.
Choosing the pull request as the deliverable forces everything upstream of it to be real. The agent has to clone the actual repository. It has to run your test command and deal with what comes back. It has to reproduce the bug before claiming to have fixed it. A snippet can be plausible; a pull request with a red suite is just wrong, in public.
What we believe
Autonomy is a question about blast radius
It's reasonable to let an agent work unattended when the worst it can produce is a pull request nobody merges. Every design decision — the throwaway sandbox, the scoped token, the fact that nothing merges itself — exists to keep that sentence true.
A question is more useful than a refusal
When a ticket is vague, the valuable move is to sharpen it, not to stop. Elisha's gate gets one chance to ask something specific. “Blocked” helps nobody; “there are two coupon code paths and I can't tell which one the customer hit” is a real contribution.
A warning that fires on everything is decoration
We treat false positives as expensive. A risk rule that flags every card on the board trains people to ignore all of them, which is strictly worse than having no rule.
Report back where you were asked
The person who reported the bug should find out it's fixed without chasing anyone — in the thread they wrote it in, in plain English, not a raw tool-call log.
Where we are
Small team, San Francisco, private beta. If Elisha sounds like something your team needs, start building — and tell us what breaks. That's the whole point of the beta.
Try it on the queue you're avoiding
It's free during the beta, and the worst case is a pull request you close.