BuildAI.
عربيEN
Start Building
Advanced

Agentic AI Engineer

From AI applications to autonomous systems.

For: An engineer with the basics, going deep on agents.

✓ Design agents with reliability budgets✓ Evaluate non-deterministic systems✓ MCP, multi-agent, and agent security
✓Choose it if: You've built AI systems and want to specialize in agents.✗Don't if: You haven't worked with LLM APIs or RAG yet.
GATES · PROVE IT

Each gate is a short assessment you must pass (70%+) to unlock the next — and every gate you pass earns its stage certificate.

A0 · 5 QTool Calling FoundationsA1 · 4 QStructured Outputs & SchemasA2 · 5 QThe Agent LoopA3 · 5 QAgentic RAGA4 · 5 QMCP & the Tool EcosystemA5 · 5 QReliability, Cost & Evaluation
See the certificates you earn →
THE JOURNEY · 8 STAGES
1
STAGE 1The Model as an EngineAfter this stage you can look at an agent design and call its cost and limits before a line of code is written.
+
4 concepts · 1 build · 1 challengeStart this stage
  1. 1The context windowControl the loop's memory6 min○
  2. 2Cost & latencyNumbers before you run7 min○
  3. 3Choosing a modelThe right model per step8 min○
  4. 4What LLMs can't doKnow when to hand it a tool6 min○
  5. 5Build — A script estimating cost and time for a 20-step agent task, across two different models.○
  6. 6Challenge — The Expensive PromptThe same result could cost a tenth — if you find where the waste is.○
2
STAGE 2Tool Calling in DepthAfter this stage you can design a production tool layer: schemas the model understands, and failures that don't break the loop.
+
4 concepts · 1 build · 1 challengeStart this stage
  1. 1Designing tool schemasTools the model understands8 min○
  2. 2The tool-call loopA sturdy loop9 min○
  3. 3Handling tool errorsFailures that don't sink the agent9 min○
  4. 4Tool securitySafe, bounded tools9 min○
  5. 5Build — Customer Support AgentA tool layer for a support agent: lookup, refund, and escalate, with validation on every call.15–20h○
  6. 6Challenge — Tool Failure CascadeOne tool went down, and the whole agent collapsed behind it.○
3
STAGE 3Agent FundamentalsAfter this stage you can write an agent loop from scratch and read its trace to see how and why it thinks.
+
3 concepts · 1 build · 1 challengeStart this stage
  1. 1The ReAct patternAn agent that reasons out loud8 min○
  2. 2Tool choice & selectionThe right tool at the right time7 min○
  3. 3Chain of thoughtCleaner loop decisions6 min○
  4. 4Build — A hand-written ReAct loop: visible reasoning, tool selection, observation, and a clear stop condition.○
  5. 5Challenge — The Over-Eager Tool CallerThe agent calls a tool for everything — even 'hello'.○
4
STAGE 4Memory & PlanningAfter this stage you can build an agent that remembers, plans, and stops right — however long the task.
+
3 concepts · 1 build · 1 challengeStart this stage
  1. 1Working memoryAn agent that keeps what matters10 min○
  2. 2Bounds & budgetAn agent that stops right8 min○
  3. 3Retrieval as memoryMemory bigger than the context13 min○
  4. 4Build — An agent with an external scratchpad, a plan updated each step, and iteration and token caps.○
  5. 5Challenge — The Agent That Never StopsThe agent circles the same steps and burns the budget.○
5
STAGE 5MCPAfter this stage you can wire an agent to any MCP server and judge the wiring's security before trusting it.
+
3 concepts · 1 build · 1 challengeStart this stage
  1. 1MCP Clients & HostsBuild a Client / Host12 min○
  2. 2MCP SecuritySafe MCP14 min○
  3. 3Injection through toolsAn agent that ignores planted orders8 min○
  4. 4Build — A host aggregating two MCP servers, with an approval gate on sensitive actions and least-privilege permissions.○
  5. 5Challenge — Schema DriftA server updated, the schema changed, and the agent still sends the old parameters.○
6
STAGE 6Advanced PatternsAfter this stage you can choose an agent architecture by measurement, and say 'a single agent is enough here' with confidence.
+
4 concepts · 1 build · 1 challengeStart this stage
  1. 1Reflection agentsHigher quality at measured cost9 min○
  2. 2Evaluator–OptimizerOutput that passes the bar9 min○
  3. 3Framework or by hand?Pick the right tool10 min○
  4. 4Parallel tool callsA faster agent at the same cost7 min○
  5. 5Build — The same task in three architectures: single agent, reflection, and evaluator–optimizer — with a quality/cost comparison table.○
  6. 6Challenge — Retry StormStacked retry layers turned one small failure into a bill and a flood of calls.○
7
STAGE 7Agent Evaluation & SecurityAfter this stage you can prove with numbers that your agent works and is secure — and re-prove it after every change.
+
3 concepts · 1 build · 1 challengeStart this stage
  1. 1Prompt injection basicsAn injection-resistant agent9 min○
  2. 2Probabilistic outputStable numbers from a variable system7 min○
  3. 3Evaluating agent trajectoriesJudge the path, not just the result12 min○
  4. 4Build — An eval harness: 20 golden tasks with expected outcomes, plus an injection attack suite run on every change.○
  5. 5Challenge — Prompt Injection 101User-supplied text made the agent ignore its instructions and do something else.○
8
STAGE 8Production AgentsAfter this stage you have a real agent running in production with evidence and numbers — proof you're an Agentic AI Engineer.
+
2 concepts · 1 build · 1 challengeStart this stage
  1. 1Agent observabilitySee inside the agent as it runs11 min○
  2. 2Agent cost engineeringCost under control in production9 min○
  3. 3Build — Research AgentThe capstone: a complete research agent — loop, memory, tools, MCP, an eval harness, and a monitoring dashboard.18–25h○
  4. 4Challenge — The $40 Agent RunAn agent run costs $40 — cut it to ≤ $4 without dropping quality.○
Browse the Knowledge Map →