Symphony

Symphony turns a written Jira ticket into reviewed code and an open merge request. Autonomous coding agents pick up sanctioned work, plan it, implement it, and hand a human a change that is ready to review.

It runs inside a BaFin- and DORA-regulated estate, where speed and auditability usually pull against each other. The interesting part of the system is keeping both.

In a regulated environment, the slow part is rarely the typing. It is traceability: knowing what changed, why, by which actor, and being able to prove it later. Symphony compresses the path from a written ticket to a reviewable change while keeping that proof intact.

Stated outcome: roughly a 30 percent faster delivery cycle on the work it covers.

Managed Kubernetes with two node pools: a stable pool for the control and reconciler loop, and an ephemeral pool for agent sessions that come and go.

Each agent session runs in an isolated container with its own toolchain. A reconciler and controller-loop pattern gives the system self-healing behaviour and audit resilience: state is observed, compared to desired state, and corrected.

State is tracked across three sources kept in agreement, so the system can recover its understanding after a restart and nothing is lost between the issue tracker, the cluster, and the database.

Repository access is gated by an allow-list. Agents can only ever touch repositories that have been explicitly sanctioned, validated against the source system before any work begins.

Work is routed across models by complexity: a lighter model orchestrates, a stronger model plans. The routing is deliberate and inspectable rather than a single model doing everything.

All inference runs through an EU-resident, zero-retention path, which is what makes autonomous agents writing production code defensible inside this regulatory class.