Compiled summary

RAG trades agency for verifiability: every answer is grounded in a retrieved span you can audit. Agent architectures invert the trade — they will act on partial information, replan, and reach conclusions no single retrieval would surface, but the audit trail is a transcript, not a citation.

The interesting systems are hybrids. Toolformer and ReAct embed retrieval as a tool the agent calls; Self-RAG lets the model decide when retrieval is needed at all. The unsolved question — see Open Questions below — is whether re-ranking helps or hurts.

Atomic concepts

• RAG — retrieval-augmented generation
• AI Agent — autonomous task-decomposing system
• ReAct — reasoning + acting interleaved
• Toolformer — self-supervised tool-use training
• Self-RAG — model-controlled retrieval
• Re-ranking — cross-encoder rescoring
• Hybrid Retrieval — missing target, search fallback

Architecture sketch

  RAG                                AGENT
  ───                                ─────
  query ─→ retrieve ─→ rerank ─→     query ─→ plan ─→ act ──┐
                       ↓                          ↑         │
                     LLM ─→ answer                └────── observe
                       ↑                                    │
                    citations                            transcript
              

📝 Detailed explanation

What is it?

An AI Agent is built around a loop: plan → act → observe → revise. Unlike a function-call pipeline, the agent's plan is generated at runtime from a goal and a tool catalog, and may be revised after every observation.

Why does it matter?

Agents trade verifiability for agency. They reach conclusions no single retrieval would surface — at the cost of an audit trail that is a transcript rather than a citation. See the topic RAG vs. Agent Architectures for the full trade-off.

How does it work?

Three execution patterns dominate: ReAct (interleave thought and action tokens), Toolformer (self-supervised training over tool tokens), and Plan-and-Execute (decouple a long-horizon plan from per-step execution).

🏗️ Architecture

  ┌──────────┐    ┌──────────┐    ┌──────────┐
  │  Plan    │ → │   Act     │ → │  Observe  │
  └──────────┘    └──────────┘    └──────────┘
        ↑                              │
        └──────── revise ──────────────┘
              

💡 Action items

• Short term — read Yao et al. 2022 (ReAct) and run the canonical hotpot QA replication.
• Long term — evaluate LangGraph vs. hand-rolled state machines for production agents.

🔗 Related concepts

• RAG — the verifiability counterpoint
• Tool Use
• Reflection
• Multi-Agent Coordination