An agent can write a JavaScript workflow for a specific task, fan work out to subagents, and orchestrate the results. The documented primitives are agent(), parallel(), pipeline(), and decide().
The decide() primitive calls a decision model, including OpenAI's Decisions API or TypeSafe's Jev, for structured choices, scales, or yes/no answers with confidence scores. Letta says workflows can assign subagents to connected models that support structured outputs, with BYOK connections including OpenAI, OpenRouter, and Together AI.
Letta also changed /init: the agent writes a workflow to distribute review of local files and past Claude Code and Codex sessions across subagents. Letta describes those subagents as stateless, with the orchestrator providing excerpts or pointers to external memory.
The announcement does not establish benchmarked quality, total workflow cost, reliability at scale, or comparative performance against other orchestration systems. Source: https://www.letta.com/blog/dynamic-workflows