TANGO: Task-Adaptive Multi-Agent Orchestration with Information routing and structured memory

Multi-agent systems can improve complex reasoning in LLMs, but many frameworks use fixed worker roles, predetermined communication, or terminal aggregation alone. We present TANGO, a task-adaptive multi-agent framework that jointly designs a query-specific portfolio of agents and orchestrates communication between them during execution. For each task, a meta-agent designs workers with complementary reasoning strategies and tool configurations. Each worker maintains a structured memory state that records its progress, hypotheses, unresolved questions, and evidence references. At periodic checkpoints, the meta-agent inspects these states, dynamically constructs a sparse directed communication topology, and selects the relevant information that needs to be routed. After execution, the meta-agent uses the workers' terminal answers and structured memories to resolve disagreements and produce the final answer. We evaluate TANGO with Gemini-2.5-Flash on GAIA, GPQA-Diamond, Omni-MATH, and HumanEval. On GAIA, TANGO reaches 55.15\% mean pass@1 and 63.03\% pass@2. TANGO further achieves strong results across GPQA-Diamond, Omni-MATH, and HumanEval, demonstrating that jointly designing agent portfolios, communication, and synthesis improves performance across diverse reasoning tasks. The code is available \href{https://anonymous.4open.science/r/TANGO-1B0B/README.md}{here}.