Cross-domain Agentic Workflow Generation
From Fragile Workflows to Composable Intelligence
A New Paradigm for Cross-Domain Agentic Systems
Date: March 26, 2026
Paper: Learning to Compose for Cross-domain Agentic Workflow Generation
The Core Problem
Agentic workflows today are fragile.
You build a system that works beautifully for one domain—say, coding—and it quickly falls apart when applied to another, like data analysis or customer support. Each domain demands its own handcrafted logic, its own tuning, its own endless cycle of debugging.
Result:
Every new use case feels like starting from zero.
Why Current Systems Struggle
Most modern approaches rely on iterative refinement:
- Generate a workflow
- Test it
- Fix it
- Repeat… (often 20+ times)
While this can eventually produce something usable, it comes with major drawbacks:
- ⏱️ High latency
- 💸 Expensive iteration cycles
- 🧩 Poor generalization across domains
Even after all that effort, workflows remain brittle and domain-specific.
A Shift in Thinking: From Iteration to Composition
The new research introduces a fundamentally different approach:
Decompose → Recompose → Decide
Instead of repeatedly fixing workflows, the system learns reusable capabilities that can be recombined for new tasks—in a single pass.
How the New Approach Works
1. Decompose
Extract reusable workflow patterns from diverse domains.
These aren’t full workflows—they’re building blocks of reasoning and execution.
2. Recompose
For a new task, assemble a sparse combination of these learned patterns.
Think of it as selecting the right tools from a well-organized toolkit rather than building tools from scratch.
3. Decide
Evaluate which capabilities contributed to success (or failure) using counterfactual analysis.
This allows the system to improve its internal representations—not by trial-and-error loops, but by structured learning.
What Makes This Breakthrough Important?
The results are striking:
- 🚀 Single-pass generation outperforms state-of-the-art systems using 20 iterations
- 🌍 Works across multi-domain, cross-domain, and unseen tasks
- ⚡ Lower latency and cost
- 🔓 Built on open-source models, reducing dependency on proprietary APIs
What This Means for Builders
This is more than an optimization—it’s a paradigm shift.
Old mindset:
Design workflows per use case
New mindset:
Learn composable capabilities that generalize
Instead of engineering solutions repeatedly, you invest in a library of transferable primitives.
Impact:
- Less per-domain engineering effort
- More robust and scalable systems
- Faster deployment across new problem spaces
Limitations to Keep in Mind
This approach isn’t magic—it comes with trade-offs:
- Requires diverse training domains to learn meaningful primitives
- Assumes tasks can be cleanly decomposed (not always true)
- Counterfactual attribution adds training complexity
The Big Takeaway
Stop designing workflows. Start learning workflow primitives.
The future of agentic systems isn’t about crafting perfect pipelines—it’s about building composable intelligence.
Teams that embrace this shift will move faster, generalize better, and ultimately outperform those stuck in domain-by-domain design loops.
Read the Paper
📄 https://arxiv.org/pdf/2602.11114
If you're building multi-agent systems, this is one of those ideas worth internalizing early.
Get Updates
New posts on systems thinking, AI, and building things. No spam, unsubscribe anytime.