Tomoko Arisawa

Repeated pipeline execution must be safe for automated recovery to work reliably.

AI agents need review tools built for pipeline speed, not human-paced workflows.

Agents can generate passing tests without specifying actual business requirements.

It's a self-reinforcing feedback loop, not a pipeline with human handoffs.

Audit agent transcripts to spot unused skills before they drain every session.

Agents can fix performance regressions faster than humans can review them.

AI agents overwhelm code review systems built for human-speed development.

AI agents need telemetry built for machines, not dashboards built for humans.

True automated rollback requires health checks, deployment history.

Agents need production failures routed back into their coding loops.

AI agents are colliding on code at rates that choke build pipelines.