Are you interested to start your career or switch your career into GenAI? After working 3 years in GenAI, I came up with this AgenticAI Roadmap.

This is good enough to make your strong career in GenAI.

Phase 1: Retrieval Foundations

1. RAG Fundamentals

1.1 Simple RAG Embed the query, retrieve the top matching chunks, and pass them to the LLM as context.

1.2 Chunking Strategies Fixed, recursive, semantic, document-aware and hierarchical splitting. How you chunk often matters more than which model you use.

1.3 Embeddings Turn text into vectors so the system can search by meaning rather than exact keywords.

1.4 Vector Databases pgvector, Qdrant, FAISS and Elasticsearch store vectors and serve fast similarity search.

1.5 Metadata Filtering Narrow retrieval by source, date, owner or type before similarity ranking.

2. Advanced RAG

2.1 Hybrid RAG Combine dense retrieval with sparse BM25 and merge the two result lists using Reciprocal Rank Fusion.

2.2 Reranking Re-score the retrieved candidates with a cross-encoder so the best chunks reach the prompt.

2.3 Multi-Query RAG Rewrite one question into several queries to widen recall.

2.4 HyDE Generate a hypothetical answer first, then retrieve using that answer's embedding.

2.5 Parent-Child RAG Match on small chunks but return their larger parent sections for fuller context.

2.6 Multi-Hop RAG Chain retrievals so each step's answer feeds the next query.

2.7 Adaptive RAG Route each query to no retrieval, single-step or multi-step retrieval based on its difficulty.

2.8 Corrective RAG (CRAG) Grade the retrieved documents and fall back to rewriting or web search when they are weak.

2.9 Self-RAG The model decides when to retrieve and critiques its own output against the sources.

2.10 GraphRAG Build a knowledge graph of entities and relations, then retrieve across it.

2.11 Agentic RAG An agent plans, picks retrieval tools and iterates until it can answer.

2.12 Multimodal RAG Retrieve over images, tables, charts and audio as well as text.

Phase 2: Engineering the System

3. Context Engineering

3.1 System Instructions The fixed role, rules and output format the model works under.

3.2 Dynamic Context Information injected at runtime, such as the user, time, state or task.

3.3 Retrieved Context Documents pulled in by RAG for the current query.

3.4 Conversation Context The prior turns kept, trimmed or summarized.

3.5 Memory Facts carried across sessions and recalled when relevant.

3.6 Tool Context Tool definitions plus the results returned by tool calls.

3.7 Context Compression Summarize or prune content so more signal fits in fewer tokens.

3.8 Context Selection Choose which candidate pieces of context make it into the prompt.

3.9 Context Prioritization Order and place content so the most important parts get the model's attention.

3.10 Context Window Management Budget tokens across all sources and stay within the model's limit.

4. Harness Engineering

4.1 Model The LLM at the core, along with its settings.

4.2 Context What the harness feeds the model on each step.

4.3 Memory State the harness persists between steps and sessions.

4.4 Tools The actions the model can call.

4.5 Permissions What each tool or agent is allowed to touch.

4.6 Guardrails Input and output checks that block unsafe or off-policy behaviour.

4.7 Execution Running tool calls safely, with sandboxing and timeouts.

4.8 Verification Checking outputs and results before they are accepted.

4.9 Retry / Recovery Handling failures, timeouts and bad outputs gracefully.

4.10 Audit Logging every decision and action for review.

5. Loop Engineering

5.1 Generate → Evaluate → Regenerate Produce a draft, score it, and improve it until it passes.

5.2 Retrieve → Evaluate → Retrieve Again Check retrieval quality and search again when the context is insufficient.

5.3 Plan → Execute → Observe → Replan The core agent loop, adjusted after each observation.

5.4 Critic / Reviewer Loop A second model or prompt reviews the work before it is accepted.

5.5 Reflection The model reasons about its own mistakes and how to fix them.

5.6 Self-Correction The model revises its output using errors or feedback.

5.7 Retry Strategy When to retry, how many times, and with what changes.

5.8 Loop Termination Stop conditions such as a quality threshold, step limit or budget so loops never run forever.

Phase 3: Agents

6. Agentic AI

6.1 Tool Calling The model emits structured calls to functions and APIs.

6.2 ReAct Interleave reasoning steps with actions and observations.

6.3 Planning Break a goal into ordered sub-tasks before acting.

6.4 Single Agent One agent with tools handles the whole task.

6.5 Multi-Agent Several specialised agents collaborate on a task.

6.6 Supervisor / Worker A supervisor delegates tasks to worker agents and merges their results.

6.7 Planner / Executor One agent plans while another carries out the steps.

6.8 Router Agent Sends each request to the right agent, tool or model.

6.9 Critic Agent Reviews other agents' outputs and flags problems.

6.10 Human-in-the-Loop A person approves or corrects the agent at key points.

6.11 Durable Agents Long-running agents that survive restarts through checkpointed state.

6.12 Agentic RAG Retrieval driven by an agent's own decisions.

7. MCP / Tool Ecosystem

7.1 MCP Architecture Agent → MCP Client → MCP Server. The server exposes tools in a standard way.

7.2 MCP Server Integrations Databases, REST APIs, search, files and enterprise systems.

7.3 A2A Agent-to-agent protocols that let independent agents talk to each other.

8. Memory

8.1 Working Memory Scratch state for the current task.

8.2 Conversation Memory What was said in this session.

8.3 Episodic Memory Records of past events and interactions.

8.4 Semantic Memory Distilled facts and knowledge.

8.5 Procedural Memory Learned how-tos and skills.

8.6 User / Profile Memory Stable facts and preferences about the user.

8.7 Long-Term Memory Persistent storage across sessions.

8.8 Memory Tooling Mem0, TencentDB memory approaches, LangGraph persistence and checkpoints, Redis, vector DB memory and graph-based memory.

8.9 Memory Lifecycle Write → extraction → consolidation → retrieval → forgetting / TTL → conflict resolution.

9. Prompt / Program Engineering

9.1 Zero-Shot Instructions only, with no examples.

9.2 Few-Shot Instructions plus a handful of worked examples.

9.3 DSPy Treat prompts as programs and optimise them automatically against a metric.

Phase 4: Model Adaptation and Evaluation

10. Fine-Tuning

10.1 Prompt Engineering Adapt the model's behaviour without changing its weights.

10.2 RAG Adapt through external knowledge instead of training.

10.3 SFT Supervised fine-tuning on input-output pairs, including instruction tuning.

10.4 PEFT Parameter-efficient tuning with LoRA and QLoRA: train small adapters instead of the full model.

10.5 Preference Optimization Align the model to preferred answers using DPO.

10.6 Reinforcement Fine-Tuning Train with reward signals from graders or verifiers.

10.7 Distillation Teach a small model to imitate a larger one.

10.8 Continued Pretraining Extend the model's knowledge with domain text.

10.9 POC: RAG vs Fine-Tuning vs RAG + Fine-Tuning Compare the three approaches on the same task and dataset.

11. Evaluation

11.1 RAG Evaluation RAGAS metrics: faithfulness, context precision, context recall and answer relevance.

11.2 LLM-as-a-Judge An LLM scores outputs, using DeepEval or a custom judge.

11.3 Custom Metrics Domain-specific scores that fit your use case.

11.4 Golden Dataset A curated set of questions with expected answers.

11.5 Regression Testing Re-run evals on every change to catch quality drops.

11.6 Human Evaluation People rate outputs as the ground truth.

11.7 Pairwise Evaluation Compare two outputs head to head.

11.8 Agent Evaluation Score task success, number of steps and trajectory quality.

11.9 Tool-Call Evaluation Check that the right tool was called with the right arguments.

11.10 POC: RAGAS vs DeepEval vs Custom LLM-as-a-Judge vs Human Labels Measure how well each automated method agrees with human judgement.

Phase 5: Production

12. Observability

12.1 Tracing Tools Langfuse, Arize Phoenix, OpenTelemetry and LangSmith.

12.2 What to Trace User query, retrieval, reranking, prompt, model, tool calls, tokens, cost, latency and evaluation score.

13. AI Safety / Guardrails

13.1 Prompt Injection Detect instructions hidden in inputs or retrieved content.

13.2 Jailbreak Detection Catch attempts to get around the model's rules.

13.3 PII Detection Find and mask personal data.

13.4 Data Leakage Prevent sensitive data from leaving through outputs or tools.

13.5 Toxicity Filter harmful or abusive language.

13.6 Hallucination Detection Flag claims the sources do not support.

13.7 Grounding Verification Check each answer against the retrieved evidence.

13.8 Tool Permissions Limit what tools and agents are allowed to do.

13.9 Content Filtering Apply policy filters to inputs and outputs.

13.10 Human Approval Require sign-off before high-risk actions.

14. Multimodal AI

14.1 Text → Image Understanding Ask questions about images.

14.2 PDF / Document Intelligence Extract structure and meaning from documents.

14.3 OCR + LLM Read scanned text, then reason over it.

14.4 Charts / Tables Interpret visual data and tabular data.

14.5 Vision-Language Models Models that take images and text together.

14.6 Audio → LLM Feed audio into the model.

14.7 Speech → LLM Transcribe speech, then process it.

14.8 Multimodal RAG Retrieve across several modalities.

15. Structured Generation / Constrained AI

15.1 JSON Schema Define the exact output shape the model must produce.

15.2 Pydantic Structured Output Typed Python models for LLM output, a natural fit with FastAPI.

15.3 Function Calling Get arguments back that match a function's signature.

15.4 Constrained Decoding Force tokens to follow a grammar or schema during generation.

15.5 Validation / Retry Validate the output and re-prompt when it fails.

15.6 Schema Repair Fix almost-valid output instead of discarding it.

15.7 Example Flow User → LLM → Pydantic schema → validation → business API.

16. Production Model Engineering

16.1 Model Routing Send each request to the best-fit model.

16.2 Model Fallback Switch to a backup model when the primary fails.

16.3 Small vs Large Models Use cheap models where they suffice and large ones where they are needed.

16.4 Quantization Shrink model weights for faster, cheaper inference.

16.5 Caching Reuse responses or prompt prefixes.

16.6 Batching Group requests to increase throughput.

16.7 Streaming Return tokens as they are generated.

16.8 Rate Limits Handle provider limits and quotas.

16.9 Token Optimization Cut tokens without losing quality.

16.10 Cost / Latency / Quality Trade-offs Balance the three for each use case.

---

POC Tracker

01 RAG (Fundamentals done) Simple, Hybrid, Reranking, Multi-Hop, HyDE, GraphRAG, Adaptive / CRAG / Self-RAG, Agentic RAG.

02 Context Engineering Selection, compression, prioritization and window management.

03 Harness Engineering Tools, permissions, guardrails, verification, recovery and audit.

04 Loop Engineering Evaluate-and-retry loops with clear termination.

05 Agentic AI Tool calling, single agent, multi-agent, human-in-the-loop.

06 MCP / A2A MCP servers and agent-to-agent communication.

07 Memory Layered memory, Mem0, TencentDB.

08 Prompt / Program Engineering DSPy.

09 Fine-Tuning SFT, LoRA / QLoRA, DPO, distillation.

10 Evaluation RAGAS, DeepEval, LLM-as-a-Judge, custom evaluation.

11 Observability Langfuse, Phoenix, OpenTelemetry.

12 Guardrails / AI Safety Injection, PII, hallucination and permission controls.

13 Multimodal AI Documents, OCR, vision, audio and speech.

14 Structured Generation Schema-validated output into business APIs.

15 Production AI Routing, caching, cost, latency, security, monitoring.