Agent architectures
Agent architectures we build
Unmatrix AI builds enterprise AI agents on three architectures: autonomous agents that decide their own path with a ReAct loop, LangGraph workflow agents that follow an explicit graph and let the model decide only at chosen nodes, and RAG agents that answer from your documents with hybrid search and citations. The right one depends on how well the work is known in advance.
Last updated
Comparison
Which architecture fits which work?
The deciding question is how well the process is known before the run starts.
| Criterion | Autonomous (ReAct) | LangGraph workflow | RAG |
|---|---|---|---|
| Who decides the path | The model, at every step | The graph. The model decides only at chosen nodes | A fixed pipeline: retrieve, then generate |
| Best for | Investigations across systems where the route varies | Known, repeatable processes with a few decision points | Answering questions from documents, with citations |
| Predictability | Lower, bounded by step limits and tool allowlists | High. The process is explicit and versioned | High. The same steps every time |
| Typical run time | Seconds to a minute | Milliseconds to seconds per node. Can pause for days | A few seconds |
| Human checkpoints | Before any side effect | Interrupt nodes wherever needed | Feedback on answers and an escalation path |
| Cost per run | Highest. Several model calls | Moderate. Model calls only at decision nodes | Lowest. One generation per question |
The three
Each architecture, in detail
- Read about autonomous agents (react)
Autonomous agents (ReAct)
Agents that reason, call a tool, read the result and decide the next step in a loop, inside a harness that enforces tool allowlists, step limits and human review.
Best for
The route to the answer depends on what each lookup returns
- Read about langgraph workflow agents
LangGraph workflow agents
Graph-based workflows where the model decides only at chosen nodes, with checkpointed state and human-in-the-loop interrupts. Example: a resume screening agent.
Best for
The process is known and repeatable, with a few decision points
- Read about rag agents
RAG agents
Document ingestion with permissions, hybrid keyword and vector retrieval with reranking, and grounded answers with citations or an honest not-found.
Best for
Staff spend time searching policies, contracts, manuals or tickets for answers that already exist
Every architecture ships with
- Read
Security and guardrails
How every Unmatrix AI system handles PII and PHI, defends against prompt injection, limits what an agent can do, and is tested before go-live.
- Read
On-prem LLMs
Deploying and operating open-source models such as Llama, Mistral and Qwen inside your data centre or private cloud for agentic workflows that cannot use hosted APIs.
- Read
Forward deployed
Two to four engineers embed at your office for twelve weeks to take one agentic workflow from scoping to production, then hand it to your team.
Questions
Choosing an architecture
Which agent architecture should we choose?
Start from the work. If the path varies case by case, an autonomous agent. If the process is known and needs to be auditable, a LangGraph workflow. If the job is answering from documents, a RAG agent. Most enterprise systems end up combining two: a workflow that calls a RAG step, or an autonomous agent with a workflow around it.
Can one system combine them?
Yes, and it usually does. A LangGraph workflow can contain a node that runs an autonomous sub-agent for an investigation step, and both can call a RAG retriever as a tool.
Do you build on LangGraph only?
LangGraph is our default for workflows. Autonomous loops are sometimes a custom harness. We choose per system, and you own the code either way.
Which models do you use?
Frontier models where reasoning quality matters, small models for extraction and classification, and open-source models on-prem for regulated environments. Every agent sits behind a gateway so the model can change without rewriting the agent.
Next step
Not sure which one you need? Bring the workflow, not the architecture.
Thirty minutes on where the work piles up. We will say which architecture fits, and whether an agent is the right tool at all.