Understanding Workflows in Solace Agent Mesh : new blog post 
We live in a deterministic world, especially our IT systems.
Which means, given a scenario or input, we EXPECT our IT systems to give back the same output; not similar or approximate but the same EXACT output.
This is especially true of our business processes which have to execute in the same flow every time every day given the same inputs.
This however clashes (sometimes horribly) with the non-deterministic behaviour of LLMs and LLM backed orchestrators.
This leads to a fundamental question : When should an AI agent system decide what to do vs. be told what to do?
Thatโs the core topic behind the Orchestrator vs. Workflow design choice in Solace Agent Mesh and one that I have explored in my new blog.
An ๐ข๐ฟ๐ฐ๐ต๐ฒ๐๐๐ฟ๐ฎ๐๐ผ๐ฟ uses an LLM to decide at runtime which agents to invoke, in what sequence, and how to assemble the result. Flexible, dynamic, powerful. But also non-deterministic by nature.
A ๐ช๐ผ๐ฟ๐ธ๐ณ๐น๐ผ๐ flips that: you define the execution graph upfront โ what runs, in what order, and what data flows between each step. Predictable. Auditable. Easier to reason about when something goes wrong.
Neither is universally better. The right answer depends on whether your problem is better solved by emergent intelligence or explicit control. If you cannot afford surprises in your execution path, choose a workflow.
This distinction mirrors a tension thatโs been in software engineering forever : declarative vs. imperative. It just looks a little different when your โfunctionsโ are AI agents.
I wrote up a deep dive on how workflows are designed in Solace Agent Mesh โ when to reach for them, how data flows through them, and what makes them different from pure LLM orchestration.
Link to the blog : https://solace.com/blog/understanding-workflows-in-solace-agent-mesh/
I would highly appreciate any feedback and thoughts on the blog and am available for a detailed discussion if you want to discuss more.