Orchestration Beats Intelligence for Reliable AI Agents

Recursive Language Models (RLMs) treat the prompt itself as a variable, merging tool calling and reasoning into a recursive, code-execution-driven loop. This approach lets a 9B model beat GPT-5 and Opus on long reasoning benchmarks by decomposing complex problems into sub-agent calls.

The real bottleneck in coding agents is not intelligence but specification, verification, and orchestration—what the speaker calls 'mismanaged geniuses'.

For engineering leaders, the implication is clear: investing in workflow design and recursive orchestration can yield greater reliability than chasing larger models. Tools like OpenProse and Claude Code’s dynamic workflows now make it practical to encode these patterns into everyday agent usage, capturing golden sessions as repeatable Prose programs.

The tradeoff is increased complexity. Recursive chains introduce new failure modes, debug overhead, and governance challenges. Teams must design verifiable task decomposition and recursion depth limits. While benchmarks are impressive, large-scale production evidence is still thin, and the drama around 'too hot to benchmark' distracts from the operational work of making agents trustworthy.