Modern agentic systems demand both flexibility and control. In our experience building cognitive automation across complex real-world environments, we found that both DAG-based architectures and SDK-first agent frameworks consistently fall short.
This is why we moved away from them entirely.
Directed Acyclic Graphs (DAGs) were a tempting early abstraction. But in practice, they became a rigid, brittle representation of business logic:
- Hard to refactor – Every change to flow logic means editing node connections and dependencies.
- Poor readability – Visual graphs look nice in theory, but quickly become unreadable in real workflows.
- Bakes logic into structure – Instead of separating behavior from architecture, DAGs entangle them.
- No state discipline – The infamous "state dict" or fragile JSON blob becomes the only way to pass context.
In the end, DAGs turn dynamic processes into static flowcharts — the exact opposite of what agentic systems should be.
On the flip side, SDK-based frameworks (especially those pushed by LLM vendors) try to abstract away the complexity — but swing too far in the other direction:
- Too much magic – Routing, tool usage, retries, context — all hidden behind black-box decorators or prompt wrappers.
- LLM-centric logic – Delegates core execution to stochastic models with no control flow guarantees.
- Difficult to audit – Hard to debug, test, or trace failures in production.
- Not composable – You can't easily extract or reuse components. You have to adopt their full stack or nothing.
These SDKs may make for great demos. But when complexity scales — they crack.
CAA (Compositional Agentic Architecture) is our answer:
- Explicit control flow (not inferred by the model)
- Layered separation between context, memory, execution, and tools
- Observable systems with typed state and versioned prompts
- Composable agents with clear responsibilities
We don’t just build agents. We build software systems that behave like agents.
- Everyone’s Building AI Agent Frameworks – Most Are Getting It Wrong
- Beyond Frameworks: Building Production-Grade Agent Systems
Agentic AI isn’t about graphs or SDKs. It’s about designing for complexity — and doing it with the discipline of software engineering.