Conventional wisdom says multi-agent intelligence requires back-and-forth discussion to converge on a solution. TopoDIM throws a wrench in that theory.
The paper introduces a “one-shot” topology generation method. Instead of forcing agents through sequential rounds of dialogue—which incurs massive latency and computation costs—TopoDIM allows agents to autonomously construct their own communication structures instantly. It decentralizes execution so agents decide how to interact—debate, evaluate, or collaborate—without a central coordinator. The data is stark: this approach reduces total token consumption by 46.41% while actually improving average performance by 1.50% over state-of-the-art models.
This validates our core thesis at MachineMachine: AI organizations that learn their own protocols are more resilient. Most builders are still manually wiring nodes—Manager routes to Coder, Coder routes to Reviewer. It is static and expensive. TopoDIM shows that when agents autonomously construct heterogeneous communication structures, they don’t just save costs; they actually perform better. The “org chart” becomes a dynamic output of the system, not a hardcoded input from the founder.
The limitation? One-shot generation is high stakes. If the initial topology is wrong, the system fails because there is no conversational recovery loop. Complex, ambiguous problems often require the “messy middle” of conversation that TopoDIM optimizes away. You cannot plan for unknown unknowns in a single pass.
We’re integrating these principles into our upcoming BenchmarkSuite v2 to handle initial context accrual more efficiently. If you want to build resilient systems that can structure themselves, join the waitlist: /early-access
MachineMachine is building the platform for autonomous AI organizations. Early access →