Symphony

Symphony is a daemon on a single host. It polls a Jira board, and when a ticket is ready it starts one process per repository, each running an autonomous coding agent against its own checkout. The agent writes the change, commits, pushes, and opens the merge request itself. Symphony never opens one and never merges. It is not a CI system either: the repository's own pipeline is what decides whether the work holds up, and a person is what puts it in production.

In an estate this size the first real decision is where the change belongs, and it is the one a model is most willing to get confidently wrong. So the model never names a repository. It proposes a group and some hints, a resolver queries the live API and ranks what actually exists, and anything it cannot confirm is dropped.

That gets it right 95 percent of the time, and the other 5 percent is the part worth building. Model confidence runs high almost every time, so it carries no information at all. The gap between the top two candidates is the only thing that says whether they are genuinely confusable, and a gap too small to trust is what puts the ticket in front of a person.

Pushing the branch starts everything else on its own. CI picks it up and builds the infrastructure plan, and nothing moves until that plan is green and the agent has read it back and signed it off. Both are mandatory, and anything short of an explicit pass is a fail. When it fails the job log is cut down to the lines that matter before it goes back to the agent, so what the agent gets is the error and not ten thousand lines around it. Three attempts, about a minute each.

The orchestrator can be killed at any moment and the design assumes it will be. On restart it reads what is on disk, re-attaches to the workers still running, and re-dispatches the ones that died. Work that already landed is never redone: a branch pushed before the crash is never pushed again, and when the state cannot be rebuilt with confidence the ticket goes to a person instead of back through the machine.

Agents run in a sandboxed workspace, and the host this runs on today is a test box where that sandbox is not enforced yet. Repository access is gated by an allow-list with three enforcement points, and scoping it to the production set is still ahead. Retry budgets live in memory, so a restart forgives a ticket that had nearly run out of attempts.