Course Objective & Scope
On long, multi-step tasks, ReAct agents that rethink at every step can become expensive and unstable. The Plan-and-Execute pattern first produces an explicit plan, runs the steps with smaller models or in parallel, and updates the plan when needed.
The course implements plan generation as structured output, step dependency graphs, the ReWOO and LLMCompiler approaches and replanning triggers. Participants solve the same task with ReAct and with Plan-and-Execute and write a report comparing quality, latency and cost.
Who Should Attend?
- AI engineers building agents for complex workflows
- Developers creating research, analysis or reporting agents
- Platform teams aiming to optimise agent cost and latency
- Prerequisite: basic tool calling and ReAct experience
Key Takeaways
- Turn tasks into verifiable, structured plans
- Build an executor that runs independent steps in parallel
- Define triggers and fallback rules for replanning
- Cut cost by choosing different models for planner and executor
- Choose between ReAct and Plan-and-Execute based on measurement
Curriculum
Why Plan First?
- Cost and consistency problems of ReAct on long tasks
- Separating planner, executor and replanner roles
Plan Generation and Validation
- Getting the plan as structured output with JSON Schema
- Step dependencies and precondition checks
- Rejecting invalid or risky plans
Execution Strategies
- Sequential execution versus DAG-based parallel execution
- ReWOO: plans with variable references solved in one pass
- An LLMCompiler-style task scheduler
Replanning and Error Handling
- Triggers: failed steps, unexpected results and new information
- Updating the plan while preserving partial results
Comparative Lab
- Solving the same task with ReAct and Plan-and-Execute
- Measuring quality, latency and token cost
- Hybrid approaches and selection criteria
Other Courses in This Category
-
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
-
Hands-on course on designing tools agents can use reliably: naming, descriptions, JSON Schema parameters, error messages, tool contracts, versioning and permission scopes.
- Live Course
- 5 modules
-
Hands-on technical course on building production-ready agent systems with LangGraph: StateGraph, conditional edges, checkpointers, interrupt-based human-in-the-loop, subgraphs and the Send API.
- Live Course
- 6 modules
-
Course on building tool-based agents with the OpenAI Agents SDK and Responses API: function tools, hosted tools such as web search and file search, handoffs, guardrails, sessions and built-in tracing.
- Live Course
- 5 modules
-
Hands-on course on building Model Context Protocol servers that expose tools, resources and prompts, and clients that consume them: stdio and Streamable HTTP, OAuth-based authorisation, security and testing.
- Live Course
- 5 modules
-
Designing multi-agent systems with supervisor, hierarchical, orchestrator-worker and swarm patterns: context isolation, task routing, cost control and a comparison of LangGraph, CrewAI and Microsoft Agent Framework.
- Live Course
- 5 modules
Get in touch
