MiroFlow open-sources graph-based research framework

MiroFlow matters because it is not just another agent demo — it is an attempt to turn deep-research agents into a configurable, benchmarkable systems layer that other teams can actually reproduce and extend.

• –The core architectural idea is an agent graph rather than a fixed chain, letting teams compose specialized agents and tools in a more flexible workflow than standard single-agent loops
• –Its heavy-reasoning mode bakes in ensemble and verifier patterns as first-class execution policies, which is a more practical framing than pretending one prompt will solve every hard research task
• –The robustness story is unusually concrete: message normalization, retries, fallback behavior, and fault isolation are all explicit parts of the framework rather than hand-wavy reliability claims
• –The benchmark claims are ambitious, especially against OpenAI Deep Research and other open frameworks, so the real test will be whether outside labs can reproduce those numbers from the public repo
• –Requiring only an OpenRouter key to get started lowers the barrier for experimentation, which could make MiroFlow a useful open baseline for teams building research agents without assembling a large proprietary stack