Thread #2 - Cross-domain Agentic Workflow Generation\nDate: 2026-03-26 (Thursday)\nPaper: Learning to Compose for Cross-domain Agentic Workflow Generation (arXiv:2602.11114)\n\n1/\nCurrent agentic workflow systems break when you change domains.\n\nNew research shows a single-pass generator that beats 20-iteration refinement baselines.\n\nHere's what builders need to know: 🧵\n\n2/\nThe problem: Building AI agent workflows today is brittle.\n\nYou design a workflow for coding tasks → it fails on data analysis.\nYou tune it for research → it breaks on customer support.\n\nEvery new domain = start from scratch.\n\n3/\nCurrent approaches: Iterative refinement.\n\nGenerate a workflow → test it → fix it → repeat 20+ times.\n\nThis is slow, expensive, and still produces unstable, domain-specific behavior.\n\n4/\nThe new approach: Decompose-Recompose-Decide.\n\nInstead of iterating, they internalize reusable "workflow capabilities" that transfer across domains.\n\nKey insight: Learn a compact set of building blocks, then compose them per-task in a single pass.\n\n5/\nHow it works:\n\nDECOMPOSE → Extract reusable workflow patterns across diverse domains\nRECOMPOSE → Map each new task to a sparse mix of these patterns\nDECIDE → Attribute success/failure to specific capabilities using counterfactual analysis\n\nOne pass. No iteration loops.\n\n6/\nResults:\n\n• 1-pass generation beats SOTA baselines using 20 iterations\n• Works across multi-domain, cross-domain, AND unseen domains\n• Substantially lower generation latency and cost\n• Built on open-source LLM (not closed API dependent)\n\n7/\nWhy this matters for builders:\n\nIf you're building multi-agent systems, this is the shift from:\n"Design workflows per use case" → "Learn composable capabilities that generalize"\n\nLess engineering per domain. More robust transfer.\n\n8/\nLimitations:\n\n• Requires diverse training domains to learn good capabilities\n• Sparse composition assumes tasks decompose cleanly (not always true)\n• Counterfactual attribution adds training complexity\n\n9/\nThe takeaway:\n\nStop designing workflows. Start learning workflow primitives.\n\nThe teams that build composable agent architectures will outpace those hand-crafting per-domain solutions.\n\nPaper: https://arxiv.org/pdf/2602.11114\n\nFollow @soren_cto for research → builder insights.