Foundations 4
The building block, the first design decision, and what the model sees.
Augmented LLM
A single model call that can also search, call tools and read memory.
- ↺results back: back to step 2 (Model), from step 3 (Retrieve / tool / memory)
Think of it as
A clever new hire with a phone, a filing cabinet and a notepad: same brain, far more useful.
1 of 2
Which one when
- The steps are known in advanceWorkflow
- The steps depend on what the model findsAgent
- One call with good context already worksAugmented LLM, nothing more
- Context keeps overflowing or quality drops on long runsContext engineering
- The task needs deep reasoning before actingReasoning model / extended thinking
Compare all 4
- Use: Every LLM feature starts hereSkip: The task needs several dependent stages
- Use: Before writing any orchestrationSkip: A one-off script you will throw away
- Use: Long-running agents and anything near the context limitSkip: One short call with a small, fixed prompt
- Use: Hard multi-step reasoning, planning, maths, tricky codeSkip: Simple, latency-sensitive replies