Your Traces Just Got a New Job: Fueling Self-Fixing Code

Arize CEO Jason Lopatecki shows observability moving from human-centric dashboards to an input layer for coding agents. The new loop: an event or schedule triggers skills that fetch traces, logs, and evals into a sandbox, where an agent investigates and files an issue or PR with evidence. The engineer’s role shifts from responder to reviewer.

Signal, available in Arize’s AX platform, runs periodically or on errors, using custom skills to pull data into a repo. The demo fixes a simple stream-cancelled error, but larger fixes still need human steering. This is not full self-healing, but it slashes cold-start investigation time.

The underlying requirement is 10x more telemetry—because agents can parse volumes humans can't. The real magic isn’t the model; it’s designing skills that format raw data into LLM-friendly files. Without that, pointing a model at traces yields little.

Engineering leaders should note the organisational pivot: budgeting for more tracing, staffing for skill development, and building a review pipeline for agent-generated changes. Autonomy is still early, but the direction is clear—observability platforms become the engine of continuous, agent-driven improvement.