Cross-Domain AI Agents: Stop Designing Workflows, Start Learning Primitives

Source: arxiv.org/pdf/2602.11114v1

---

Current agentic workflow systems break when you change domains.

New research shows a single-pass generator that beats 20-iteration refinement baselines. Here's what builders need to know.

The Problem: Brittle by Design

Building AI agent workflows today is fragile:

• You design a workflow for coding tasks → it fails on data analysis
• You tune it for research → it breaks on customer support
• Every new domain = start from scratch

Current approaches rely on iterative refinement: generate a workflow → test it → fix it → repeat 20+ times. This is slow, expensive, and still produces unstable, domain-specific behavior.

The New Approach: Decompose-Recompose-Decide

Instead of iterating, the new approach internalizes reusable "workflow capabilities" that transfer across domains.

Key insight: Learn a compact set of building blocks, then compose them per-task in a single pass.

How it works

1. DECOMPOSE — Extract reusable workflow patterns across diverse domains
2. RECOMPOSE — Map each new task to a sparse mix of these patterns
3. DECIDE — Attribute success/failure to specific capabilities using counterfactual analysis

One pass. No iteration loops.

Results

• 1-pass generation beats SOTA baselines using 20 iterations
• Works across multi-domain, cross-domain, AND unseen domains
• Substantially lower generation latency and cost
• Built on open-source LLM (not closed API dependent)

Why This Matters for Builders

If you're building multi-agent systems, this is the shift from:

"Design workflows per use case" → "Learn composable capabilities that generalize"

Less engineering per domain. More robust transfer.

Limitations

• Requires diverse training domains to learn good capabilities
• Sparse composition assumes tasks decompose cleanly (not always true)
• Counterfactual attribution adds training complexity

The Takeaway

Stop designing workflows. Start learning workflow primitives.

The teams that build composable agent architectures will outpace those hand-crafting per-domain solutions.

---

Paper: https://arxiv.org/pdf/2602.11114v1