Prerequisites Basic computer literacy and a willingness to document experiments.
Learning objectives Explain the stage-specific model and its limits Apply the listed concepts using a controlled exercise Record results, failure cases and a reproducible evidence trail
Topics to learn Deterministic workflow vs agentic control Task decomposition and success conditions State, events and stopping criteria Failure containment and human approval points AI Agents specialization: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Tools Python State diagrams JSON schemas
Hands-on exercise Model a simple research or support workflow as a state machine before adding an LLM. Context for this route: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Suggested project An agent design brief with explicit scope, states and stop conditions. The outcome must directly demonstrate AI Agents, not merely list tools.
Learning resources Completion criteria The workflow has a testable objective and clearly bounded actions. Evidence must be reproducible and tied to AI Agents.
Next: Tool interfaces and structured outputs
✓ Mark this node as complete
Prerequisites Completion of the previous stage and a working local practice environment.
Learning objectives Explain the stage-specific model and its limits Apply the listed concepts using a controlled exercise Record results, failure cases and a reproducible evidence trail
Topics to learn Function/tool schemas and validation Structured JSON output and parsing Authentication and least-privilege credentials Timeouts, retries and idempotency AI Agents specialization: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Tools OpenAI-compatible API client Pydantic or JSON Schema Mock tool server
Hands-on exercise Create two read-only tools and reject malformed or unauthorised tool arguments. Context for this route: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Suggested project A schema-validated tool adapter with mock tests. The outcome must directly demonstrate AI Agents, not merely list tools.
Learning resources Completion criteria Invalid requests fail closed and tool permissions are minimal. Evidence must be reproducible and tied to AI Agents.
Next: Planning, orchestration and control flow
✓ Mark this node as complete
Prerequisites Completion of the previous stage and a working local practice environment.
Learning objectives Explain the stage-specific model and its limits Apply the listed concepts using a controlled exercise Record results, failure cases and a reproducible evidence trail
Topics to learn Plan/execute patterns and routing Conditional branches and task queues Iteration limits and budget controls Deterministic orchestration around model decisions AI Agents specialization: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Tools LangGraph or equivalent state graph Python Structured logging
Hands-on exercise Implement a bounded multi-step workflow with a maximum step count and clear exit path. Context for this route: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Suggested project A traceable orchestration graph with unit tests for each transition. The outcome must directly demonstrate AI Agents, not merely list tools.
Learning resources Completion criteria Every path has a stop condition and failure state. Evidence must be reproducible and tied to AI Agents.
Next: Memory, retrieval and context management
✓ Mark this node as complete
Prerequisites Completion of the previous stage and a working local practice environment.
Learning objectives Explain the stage-specific model and its limits Apply the listed concepts using a controlled exercise Record results, failure cases and a reproducible evidence trail
Topics to learn Conversation state vs persistent memory Retrieval indexes and provenance Context selection and summarization Prompt injection boundaries for retrieved content AI Agents specialization: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Tools Vector database or local index RAG pipeline Evaluation dataset
Hands-on exercise Build a retrieval-backed assistant that cites source records and ignores embedded instructions. Context for this route: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Suggested project A provenance-aware knowledge assistant with retrieval test cases. The outcome must directly demonstrate AI Agents, not merely list tools.
Learning resources Completion criteria Sources remain traceable and untrusted retrieved text cannot authorise actions. Evidence must be reproducible and tied to AI Agents.
Next: Safety, permissions and human-in-the-loop
✓ Mark this node as complete
Prerequisites Completion of the previous stage and a working local practice environment.
Learning objectives Explain the stage-specific model and its limits Apply the listed concepts using a controlled exercise Record results, failure cases and a reproducible evidence trail
Topics to learn Tool allowlists and capability boundaries Input/output validation and content filtering Human approval for external side effects Secrets, personal data and audit trail AI Agents specialization: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Tools Policy layer Secret manager Audit logs
Hands-on exercise Simulate a risky requested action and verify that it pauses for approval. Context for this route: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Suggested project A threat model and guarded action pipeline with negative tests. The outcome must directly demonstrate AI Agents, not merely list tools.
Learning resources Completion criteria High-impact actions cannot run without explicit authorisation. Evidence must be reproducible and tied to AI Agents.
Next: Evaluation, tracing and reliability
✓ Mark this node as complete
Prerequisites Completion of the previous stage and a working local practice environment.
Learning objectives Explain the stage-specific model and its limits Apply the listed concepts using a controlled exercise Record results, failure cases and a reproducible evidence trail
Topics to learn Task-level success metrics and golden sets Tool-call validity and trajectory review Latency, cost and retry budgets Regression tests and adversarial cases AI Agents specialization: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Tools OpenTelemetry concepts LLM evaluation harness Structured traces
Hands-on exercise Run a fixed evaluation set and label failure causes by planning, retrieval or tool execution. Context for this route: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Suggested project An evaluation report and trace viewer for a bounded workflow. The outcome must directly demonstrate AI Agents, not merely list tools.
Learning resources Completion criteria Quality and cost are measured across repeated runs, not cherry-picked examples. Evidence must be reproducible and tied to AI Agents.
Next: Deployment and operational governance
✓ Mark this node as complete
Prerequisites Completion of the previous stage and a working local practice environment.
Learning objectives Explain the stage-specific model and its limits Apply the listed concepts using a controlled exercise Record results, failure cases and a reproducible evidence trail
Topics to learn Worker isolation and queue design Rate limits and tenancy boundaries Observability and rollback Versioning prompts, tools and models AI Agents specialization: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Tools Docker Queue or worker Monitoring
Hands-on exercise Deploy a staging worker with disabled external side effects and exercise rollback. Context for this route: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Suggested project A staging deployment with an operations runbook and permission matrix. The outcome must directly demonstrate AI Agents, not merely list tools.
Learning resources Completion criteria The service can be stopped, audited and rolled back predictably. Evidence must be reproducible and tied to AI Agents.
Next: Capstone: governed AI agent
✓ Mark this node as complete
Prerequisites Completion of the previous stage and a working local practice environment.
Learning objectives Explain the stage-specific model and its limits Apply the listed concepts using a controlled exercise Record results, failure cases and a reproducible evidence trail
Topics to learn Requirements and scope review Tool contracts and state orchestration Safety, evaluation and approvals Handover and continuous evaluation AI Agents specialization: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Tools Python State graph Tool schema validation Git
Hands-on exercise Build an agent for one bounded workflow and run failure/abuse tests. Context for this route: AI Agents: focus on state machines, tool schemas, planning loops, approval gates and traceable execution.
Suggested project A complete project with source, test set, traces, threat model and deployment guide. The outcome must directly demonstrate AI Agents, not merely list tools.
Learning resources Completion criteria The system demonstrates measurable utility while preventing unauthorised side effects. Evidence must be reproducible and tied to AI Agents.
End of roadmap: Complete the final project and build your portfolio.
✓ Mark this node as complete