Meta reports building a 1,300-line orchestrator that routes more than 20 workflows to 14 specialized agents, and says the orchestrator performs zero domain work. It describes an earlier single-agent approach that became unreliable as it absorbed configuration schemas, code-generation rules, routing logic and validation requirements. Mechanism: Deterministic sequencing, field gathering and reliability gates remain in the orchestrator; domain-specific work is delegated to specialists with isolated context, and workflow templates define phase order while reusing the same agent pool.
Why now: Meta frames this as a response to production failures in a multi-repository, multi-language pipeline, rather than a prompt tweak or larger context window. The operational claim is specific: preparation for a new data-column configuration reportedly fell from three to four days by hand to under an hour, and about two-thirds of system changes land without significant human edits. First implication: For teams whose agents cross heterogeneous domains, the case makes a testable argument for keeping routing and reliability checks deterministic and moving domain context to bounded specialists.
Evidence boundary: These are Meta’s own reported internal results, not an independently evaluated benchmark; the post does not establish that the same gains transfer to other workloads. First watch condition: Watch for independent replications reporting task scope, failure rates, human-review effort and comparable before-and-after timing, not just agent counts or architecture diagrams.
The record is distinct because no prior New Runtime record was supplied for this specific production case; its reported workflow timings and reusable-template mechanism support an independent analytical entry.