Learning Path: AI Agentic Engineer
This is the core path for engineers who want to design, build, and operate production AI agents — systems that use tools, retrieve information, execute multi-step plans, and run reliably in enterprise environments.
Who this is for
- software engineers with 2+ years of backend or full-stack experience
- developers who have used LLM APIs and want to go further
- engineers tasked with building internal AI products
- anyone preparing for an AI engineering role in 2025–2026
What you will be able to build by the end
- a provider-agnostic LLM gateway with retry, tracing, and structured outputs
- a RAG pipeline with local evals and citation quality measurement
- a multi-step agent orchestrator with approval states and inspectable traces
- a safe enterprise tool layer using MCP-style patterns
- an operator-facing copilot UI with real-time agent state
- a production hardening plan covering security, evals, observability, and deployment
The path
Phase 1: LLM runtime and provider abstractions (1–2 weeks)
Before building agents, you need a solid model integration layer.
| Topic | Resource |
|---|---|
| LLM Foundations Overview | Overview |
| Prompt Contracts | Prompt Contracts |
| Streaming and Structured Outputs | Streaming and Structured Output |
Build: P01 — AI Provider Gateway
Outcome: a TypeScript module that normalises calls to OpenAI, Anthropic, and a local model behind a single interface. Add streaming and structured JSON output. Run npm run smoke.
Key concepts to internalize:
- why provider abstraction matters (vendor lock-in, cost routing, fallback)
- how streaming works at the HTTP and SDK level
- how
zodschema validation locks model output to a contract
Phase 2: Retrieval and grounded answers (2 weeks)
Agents that only use parametric knowledge hallucinate. RAG grounds them.
| Topic | Resource |
|---|---|
| RAG Overview | Overview |
| Chunking, Embeddings, Vector Search | Chunking and Embeddings |
| Hybrid Search, Reranking, Citations | Hybrid Search |
Build: P02 — Enterprise RAG Copilot
Outcome: a document retrieval pipeline with chunking, embedding, vector search, reranking, and citation output. Run npm run eval to measure retrieval quality.
Key concepts to internalize:
- chunk size trade-offs (granularity vs context)
- why BM25 + vector hybrid beats pure vector search on most enterprise datasets
- citation accuracy as a first-class eval metric
Phase 3: Agent orchestration (2 weeks)
Single LLM calls are not agents. Agents plan, use tools, retry, and ask for help.
| Topic | Resource |
|---|---|
| Agent Frameworks Overview | Overview |
| LangGraph State Machines | LangGraph State Machines |
| CrewAI, OpenAI ADK, and Multi-Agent | CrewAI and OpenAI ADK |
Build: P03 — Agent Workflow Orchestrator
Outcome: a LangGraph-based workflow with planning, tool execution, retry logic, human approval nodes, and trace output. Run npm run eval.
Key concepts to internalize:
- state machine design for agent workflows
- how to model approval checkpoints as graph edges
- why inspectable traces are not optional in enterprise agents
Phase 4: Tool calling and MCP (1 week)
Safe tool use is the difference between a useful agent and a dangerous one.
| Topic | Resource |
|---|---|
| Tool Calling Patterns | Tool Calling Patterns |
| MCP Server and Client Basics | MCP Basics |
Build: P04 — MCP Enterprise Toolkit
Outcome: an MCP-style tool server with read-only enforcement, parameter validation, audit logging, and safe error surfaces. Run npm run eval.
Key concepts to internalize:
- why read-only by default is the right starting policy
- how MCP separates resource description from execution
- audit logging as a compliance requirement, not an afterthought
Phase 5: Agentic UI (1 week)
Agents need operator interfaces. A terminal is not a product.
| Topic | Resource |
|---|---|
| Agentic UI Overview | Overview |
| AG-UI Event Streams | AG-UI Protocol |
| Angular Copilot UX Patterns | UX Patterns |
Build: P05 — Angular Agentic Copilot
Outcome: an Angular copilot shell that consumes agent event streams, renders plan states, and handles approval flows.
Phase 6: Security (1 week)
Production agents face adversarial inputs, over-privileged tool access, and regulatory scrutiny.
| Topic | Resource |
|---|---|
| Guardrails, Permissions, Approvals | Guardrails |
| Prompt Injection Threat Model | Prompt Injection |
| Enterprise Readiness Checklist | Checklist |
Build: Add input validation, a deny-list guardrail, and rate limiting to P01 or P03. Document the threat model for your orchestrator.
Phase 7: Evals and observability (1 week)
You cannot improve what you cannot measure.
| Topic | Resource |
|---|---|
| Agent Evaluation Methods | Eval Methods |
| Traces, Metrics, and Costs | Observability |
Build: Add LangSmith or OpenTelemetry tracing to P03. Write an eval that measures tool call accuracy and plan completion rate.
Phase 8: Deployment (1 week)
| Topic | Resource |
|---|---|
| Docker and CI/CD | Docker and CI |
| AWS Free Tier Path | AWS Path |
Build: Containerise P02 or P03. Add a GitHub Actions workflow that runs the eval suite on each push and posts results as a check.
Completion criteria
You are done when you can demonstrate and explain:
- a production provider gateway with structured outputs and retry
- a grounded RAG pipeline with measurable citation quality
- a multi-step orchestrator with approval states and trace output
- a safe MCP tool layer with audit logging
- an agent UI with real-time state
- at least one eval suite measuring output quality
- a deployment pipeline that runs evals in CI
Timeline
Focused learner (2–3 hours per day): 8–12 weeks
Part-time learner (5–7 hours per week): 16–20 weeks
Portfolio output
By the end of this path you have 4–6 GitHub repos that demonstrate production-grade AI engineering across the full stack. Write a brief case study for each. This is the portfolio that gets AI engineering interviews.
See Portfolio Strategy for how to present this work.