Course Objective & Scope
In multi-domain assistants, an agent can enlist another specialist in two fundamentally different ways: transferring conversation control entirely (handoff) or invoking the specialist like a function and keeping ownership of the user conversation (agent-as-tool). Choosing poorly either strands the user in the wrong specialist or wastes tokens re-summarising answers.
Using the OpenAI Agents SDK and LangGraph Command/subgraph primitives, this course implements both patterns along with input/output filtering, history scoping, return-to-triage rules and ping-pong loop prevention. Participants design and code a multi-specialist customer or operations assistant for their organisation.
Who Should Attend?
- AI engineers building multi-department or multi-specialist assistants
- Software developers creating conversational customer service and internal support bots
- Solution architects standardising enterprise assistant architectures
- Prerequisite: Python or TypeScript and basic tool calling experience
Key Takeaways
- Contrast handoffs and agent-as-tool across control flow, latency and context isolation
- Build reliable triage routing, escalation and return-of-control rules between specialists
- Write input filters that pass only scoped context payloads during delegation
- Detect and prevent infinite ping-pong handoff loops between agents
- Combine both patterns smoothly behind a unified conversational user experience
Curriculum
Delegation Architecture: Handoff vs Agent-as-Tool
- Control transfer (peer-to-peer handoff) versus centralised invocation (sub-agent as tool)
- Trade-offs in multi-turn dialogue ownership, latency and token usage
Handoff Design and Triage Routing
- Handoff tool schemas, transfer reasons and precondition checks
- Tracking the active speaker and returning control to triage after task completion
- Preventing ping-pong loops and enforcing transfer budgets
Modular Specialist Invocation with Agent-as-Tool
- Wrapping a sub-agent as a stateless, isolated function tool
- Running specialist agents in parallel and synthesising responses in the parent agent
Context Scoping and Isolation
- Full message history sharing versus structured handoff payloads
- Filtering sub-agent internal tool calls and scratchpads out of the main thread
Workshop: Enterprise Assistant with Hybrid Delegation
- Coding a multi-specialist assistant combining handoffs and agent-as-tool
- Testing routing accuracy and multi-turn conversation continuity
Other Courses in This Category
-
Course on designing reusable skill packages, progressive context disclosure, bundled scripts/references and dynamic capability loading so agents scale without overflowing their context window.
- Live Course
- 5 modules
-
Course on cross-framework and cross-vendor agent collaboration using the A2A (Agent2Agent) protocol: Agent Cards, task lifecycle, streaming updates, enterprise authentication and MCP–A2A complementarity.
- Live Course
- 5 modules
-
Training on pausing agents for human approval on high-risk tool calls and low-confidence decisions, editing tool arguments, routing asynchronous escalations via Slack/Teams/web UIs and maintaining audit trails.
- Live Course
- 5 modules
-
Hands-on training for developers on building ReAct agents from scratch: the Thought–Action–Observation loop, native tool calling, stop conditions, iteration limits and failure analysis.
- Live Course
- 5 modules
-
Technical course on agents that split multi-step tasks into planner, executor and replanner roles, covering ReWOO and LLMCompiler-style parallel task graphs, plan validation and cost optimisation.
- Live Course
- 5 modules
-
Course for teams building long-running agents: state schemas, reducers, checkpoints, thread management and state-machine workflows, illustrated with LangGraph and workflow engine examples.
- Live Course
- 5 modules
Get in touch
