Types of Agents
Wrong agent type wastes effort — a FAQ bot does not need a 12-node LangGraph.
Bicycle for short trips, truck for hauling, fleet for logistics.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Scroll inside the frame to explore · use + / − to zoom up to 200%
Agent Type Selection
Scroll inside the frame to explore · use + / − to zoom up to 200%
Match complexity to agent type — avoid over-engineering.
Key Takeaways
- 1.Reactive: input LLM output no tools.
- 2.Conversational: chat plus memory buffer.
- 3.Task-oriented: tool loop with stop rules.
- 4.Deliberative: planner then executor.
- 5.Multi-agent: supervisor plus workers.
Real Example
Scenario
Internal HR policy FAQ = conversational (buffer memory, no tools). CI PR reviewer = task-oriented ReAct (read diff → run linter → post comment). Quarterly board pack = deliberative (plan all sections, then execute).
What you would do
HR FAQ: wrong to use a supervisor graph — static RAG + chat memory is enough. PR reviewer: task loop stops when the comment is posted. Board pack: deliberative because steps are known and sequential. Monitor: p95 latency on simple FAQs (signals over-engineering).
Practice Task
Classify three features on your backlog as reactive, conversational, task-oriented, deliberative, or multi-agent. Justify each in one sentence.
Code Walkthrough
Highlighted lines show where Types of Agents happens in the code.
1def pick_agent_type(task: dict) -> str: # define a reusable function2 if task.get("single_turn") and not task.get("tools"):3 return "reactive" # return the result4 if not task.get("tools"):5 return "conversational" # return the result6 if task.get("known_steps"):7 return "deliberative" # return the result8 if task.get("scope") == "large":9 return "multi-agent" # return the result10 return "task-oriented" # return the result11
12print(pick_agent_type({"tools": True, "known_steps": True})) # deliberative13print(pick_agent_type({"tools": True, "single_turn": False})) # task-orientedCheat Sheet
Quick recap
quick ref- •Match type to task complexity
- •Reactive = no tools
- •Task-oriented = tool loop
- •Deliberative = plan first
- •Hybrid architectures common
Common Mistakes
- ✕Skipping evaluation for Types of Agents before production
- ✕No logging or tracing around types of agents steps
- ✕Ignoring cost and latency implications