Visual companion for the episode: a single strong LLM is turned into a specialist team with distinct tool-use policies — text, code, search, hybrids — then iteratively refined under roughly cost-matched inference budgets.
Switch between a single-agent baseline and the TUMIX multi-agent orchestration. Hover nodes and links for local details.
Agents differ less by weights than by permissions, prompting, and willingness to call tools. Explore task classes and the expected utility of each pathway.
Bars show benchmark performance from the episode’s cited figures. Toggle model family and compare accuracy, token cost, and latency pressure.
This map situates TUMIX among self-consistency, tool-using agents, executable reasoning, and agent-orchestration systems.
Primary citations and neighboring work mentioned in the episode.