Course Objective & Scope
When a single agent is overloaded with too many tools and responsibilities, tool selection degrades and instructions clash. This course teaches when to move from a single agent to a multi-agent architecture and how to coordinate specialised agents in practice.
Participants build supervisor–worker, hierarchical, sequential handoff and parallel research patterns with LangGraph and the OpenAI Agents SDK, examine Agent-to-Agent (A2A) communication principles, and deliver a working multi-agent example and architecture checklist on a case tailored to their organisation.
Who Should Attend?
- AI and software engineers building complex agent systems
- Solution and system architects designing enterprise agent platforms
- R&D and AI engineering leads
- Prerequisite: Python and single-agent development experience
Key Takeaways
- Identify when to transition from a single agent to a multi-agent architecture
- Code supervisor, handoff and hierarchical orchestration patterns
- Design inter-agent messaging schemas and shared state spaces
- Prevent deadlocks, cost explosions and cascading errors in multi-agent loops
- Build a reference multi-agent application suited to an enterprise scenario
Curriculum
When Multi-Agent? Architectural Decision Criteria
- Single-agent limits: tool overload and context bloat
- Benefits of specialisation, isolation and parallel execution
Orchestration Patterns
- Supervisor–worker, hierarchical trees and peer-to-peer handoffs
- Parallel fan-out/fan-in and debate/judge patterns
Framework Comparison and Implementation
- LangGraph subgraphs, OpenAI Agents SDK handoffs, CrewAI and AutoGen
- State sharing: shared scratchpad versus message passing
The A2A Protocol and Interoperability
- Agent Card definitions, capability discovery and task lifecycle
- Secure communication across agents built by different teams or systems
Reliability, Cost and Observability
- Stopping handoff loops and cascading failures
- Per-agent latency and token analysis with distributed tracing
Other Courses in This Category
-
Human-in-the-loop agent architecture that preserves human oversight on critical actions: interrupt/resume flows, risk-based approval thresholds, escalation queues and agents that improve from human feedback.
- Live Course
- 5 modules
-
Connecting enterprise databases, APIs and internal systems to LLMs and agents cleanly and securely with Model Context Protocol (MCP): MCP server/client development, authorisation and enterprise integration architecture.
- Live Course
- 5 modules
-
Turning institutional knowledge, procedures and tools into modular, reusable agent skills: SKILL.md structure, progressive disclosure, templates, scripts and enterprise skill library governance.
- Live Course
- 5 modules
-
End-to-end and step-level evaluation of multi-step agent systems: trajectory evals, tool calling accuracy, task completion metrics, simulated user/environment testing and CI/CD eval pipelines.
- Live Course
- 5 modules
-
Securing tool-using agents that act on external systems: indirect prompt injection defence, delegated user authorisation with OAuth/OBO, sandbox isolation, guardrails and OWASP Agentic security controls.
- Live Course
- 5 modules
-
Foundational programme on how LLM-based agents work: model, tools, memory and the control loop; how agents differ from chatbots and workflow automation, and how to pick the right first use case for your organisation.
- Live Course
- 6 modules
Get in touch
