Become an AI agent builder: agents that use tools and finish multi-step work on their own.

When an agent is the right design, and when a workflow is better.

1. Workflows vs agents
2. The agent loop
3. Autonomy levels
4. Where agents fail
5. Cost and latency realities

Design tools models can use correctly.

1. Function calling
2. Tool descriptions and schemas
3. Error messages models can act on
4. Tool selection at scale
5. Sandboxed execution

Know the major frameworks and their trade-offs.

1. Graph-based: LangGraph
2. SDK-native: OpenAI and Claude Agent SDKs
3. Pydantic AI and Google ADK
4. When to build your own
5. Avoiding lock-in

Agents that handle long, branching tasks.

1. State machines and graphs
2. Planning and re-planning
3. Checkpoints and resumability
4. Human approval interrupts
5. Timeouts and budgets

Give agents the right memory without bloating context.

1. Short-term vs long-term memory
2. Summarisation and compaction
3. Retrieval as memory
4. User and session memory
5. Forgetting and privacy

Connect agents to tools and to each other through open protocols.

1. Model Context Protocol in depth
2. Building and securing MCP servers
3. Agent-to-agent (A2A) basics
4. Discovering capabilities
5. Protocol security

Split work across specialised agents, when it helps.

1. Supervisor and handoff patterns
2. Parallel agents
3. Shared state and conflicts
4. Debugging multi-agent runs
5. When one agent is better

Measure trajectories, not just final answers.

1. Task success metrics
2. Trajectory and tool-use grading
3. Simulated users and environments
4. Regression suites
5. Tracing and observability

Stop agents from doing damage.

1. Prompt injection through tools
2. Least-privilege permissions
3. Approvals for risky actions
4. Audit logs
5. Red-teaming your agent

Design, build, evaluate and deploy one real agent.

1. Picking a valuable task
2. Architecture and tools
3. Eval suite
4. Deployment and monitoring
5. Write-up and demo