Agentic Systems

"Let's call these agentic systems with various degrees of autonomy instead of arguing if something is or isn't an agent" by Andrew Ng in LangChain Interrupt in May 2025

Agentic Systems have three core capabilities and it actively generate queries, select tools, and retain information.

• Retrieval: Access external information

• Tools: Use services and APIs

• Memory: Remember across interactions

⚖️ LLM Calls vs Agents

LLM Calls: LLMs and tools follow predefined paths

• Predictable and consistent

• Best for defined tasks

• Lower risk

Agents: LLMs control their own processes

• Flexible decisions

• Higher autonomy

• More complex

🎨 Agentic Patterns

See Anthropic's guide for details.

• Prompt Chaining: Sequential steps. Each LLM processes previous outputs.

• Routing: Classify inputs. Direct to specialised tasks.

• Parallelisation: LLMs work simultaneously. Aggregate outputs.

• Orchestrator-Workers: Central LLM delegates to workers.

• Evaluator-Optimiser: One LLM generates. Another evaluates.

🏆 Best Practices

Tool is important

Tools need prompt engineering attention.

Iterate

1. Start Simple: Single LLM calls with retrieval

2. Test: Measure performance

3. Add Workflows: When proven beneficial

4. Consider Agents: When simpler fails ‼️

Agent Frameworks or Not?

• Start Direct: Use LLM APIs first

• Understand Code: Know what frameworks do

• Reduce Layers: Move to basic components

• Keep Control: Debug and modify freely

Build the right system for your needs. Not the most sophisticated one.