Agentic Workflows: Design, Build, and Deploy Enterprise AI Agents (online)

Course aims

By the end of the course, you will be able to:

• Map enterprise workflows, roles, tasks, and decision points into agentic system designs.
• Explain how LLMs, tools, retrieval, memory, planning, reflection, specialised agent roles, and workflow state contribute to agentic workflows.
• Distinguish between agentic demos, prototypes, and production-ready systems.
• Explain how agent identity, agent harnesses, structured skills, and context boundaries contribute to reliable enterprise agentic workflows.
• Design agentic workflows with appropriate retrieval, tool use, evaluation, security, monitoring, governance, cost control, and human oversight layer. 
• Apply secure tool-use principles, including permission boundaries, input validation, sandboxing, prompt-injection safeguards, and failure handling.
• Compare common workflow and multi agent patterns, including supervisor, hierarchical, role based, graph based, and peer to peer approaches.
• Use coding agents, structured skills, notebooks, and guided implementation patterns to support AI-assisted development.
• Define appropriate metrics, traces, audit trails, escalation paths, and observability requirements for agentic systems.
• Communicate agentic workflow designs clearly to both technical and strategic stakeholders.

Level and demands

This course is designed for strategic and technical leaders, AI-assisted development professionals, including Solution Architects, Staff Engineers, AI Leads, Executive Advisors, Executive Tech Leads, Principal Consultants, Business Strategists, technically oriented product leaders, and professionals responsible for evaluating or guiding enterprise AI systems.   

Participants may not all want to become full-time agent developers. However, they will want to acquire enough implementation literacy to guide teams, evaluate architectures, understand tradeoffs, assess tooling choices, and judge whether an agentic workflow is moving toward production readiness. 

The course is also suitable for engineers and architects who want to go deeper into implementation. These participants can use the accompanying book, code examples, notebooks, skills, and repository materials for further technical depth. 

Prerequisite knowledge expected  

We expect participants to have knowledge of enterprise systems and an overall understanding of the software development lifecycle.    

However, no advanced programming or software engineering background is required. You do not need to be a full-time software developer. You should be comfortable engaging with practical, guided technical material, including structured skills, notebooks, AI-assisted development workflows, and implementation examples.    

Anyone with an engineering or technical background can go deeper into implementation details through the book and repository materials. 

Programme details

Session 1: From LLMs to agentic workflows

This session introduces the foundations of agentic workflows and explains how large language models become part of dynamic systems.
 
You will examine the shift from single-prompt interactions to agentic systems that use state, tools, context, planning and workflow transitions. You will learn how state influences an LLM-based system’s behaviour and how transitions can be used to guide it through a structured workflow.

You will also learn how enterprise workflows can be translated into tasks, roles, decision points, agent responsibilities, tool dependencies and human checkpoints. The session concludes by exploring the key differences between demonstrations, prototypes and systems designed for real-world production.

Session 2: Agent architectures, planning, and multi agent workflows

This session focuses on how agents reason, plan, act, reflect and coordinate.

You will examine structured reasoning and action patterns, including planning, reflection, iterative improvement and test-time reasoning. You will also consider where human oversight is needed, how effective checkpoints can be built into workflows and how agent autonomy can be safely bounded in enterprise settings.

The session also compares common multi-agent architectures, including supervisor-led, hierarchical, role-based, graph-based, peer-to-peer and swarm-style approaches. You will explore how each model works, where it is most effective and how to choose the right architecture for a particular workflow.

Session 3: Models, contracts, tools, and agent harnesses

This session explores how model capabilities shape the design of reliable, production-oriented agents.

You will consider how model architecture, reasoning capabilities and the choice between open and closed models influence agentic workflows. You will also examine mixture-of-experts models and the practical optimisation trade-offs involved in selecting the right approach for different tasks and enterprise settings.

The session then focuses on the foundations of reliable execution, including structured outputs, validation, tool contracts, data flow, MCP-style interfaces and effective handoffs between agents, tools and systems.

You will also explore the agent harness as the operating layer around the model, defining an agent’s identity, role, responsibilities, authority, context boundaries, reusable skills and execution constraints.

Session 4: Secure execution, deployment, and operational controls

This session examines the production layer around agentic systems and the controls required for reliable enterprise use.

You will explore tool governance, permission boundaries, sandboxing, runtime isolation, and secure programmatic tool calling, including why prompts alone do not provide containment.

The session then considers what changes when an agentic workflow moves from a demonstration or minimum viable product into production. You will explore infrastructure and deployment requirements, as well as monitoring, fallbacks, rollback paths and escalation processes.

You will also discover how practical factors such as caching, cold starts, inference backends, tool budgets and cost controls affect the performance and operation of multi-step agentic systems.

Session 5: Evaluation, observability, and production feedback

This session focuses on how agentic systems are evaluated before and after deployment.

You will learn how to define evaluation criteria, conduct behavioural stress testing and role-play testing, analyse failure modes, and introduce early threat testing during development.

The session explores traces, metrics, logs, production monitoring and post-deployment evaluation. You will consider how production traces can be turned into custom benchmarks, how long-horizon reasoning and complex workflows can be evaluated, and how results can be communicated to technical and strategic stakeholders.

You will also consider evaluation as both a technical discipline and a governance mechanism, helping organisations improve system performance, manage risk and make informed decisions about agentic systems.

Session 6: Memory, cost, threat modeling, and production readiness

This session explores the longer-term design, cost and security considerations involved in running agentic workflows in production.

You will examine how different forms of memory, including short-term, long-term, episodic, semantic and procedural memory, support agent behaviour. You will also consider memory hygiene, context management, memory topology and the boundaries that determine what information is retained and reused.

The session also covers agentic cost multipliers, efficient system design, tool budget boundaries, state transition reuse, caching and the relationship between architecture and operational cost.

You will then explore threat modelling for AI agents, layered defences, red teaming and system-level safeguards, before completing a final production-readiness review of your capstone work.

IT requirements

This course is delivered online using Microsoft Teams. You will be required to follow and implement the instructions we send you to fully access Microsoft Teams on the University of Oxford’s secure IT network.

To participate you must be familiar with using a computer for purposes such as sending email and searching the Internet. You will also need regular access to the Internet and a computer meeting our recommended minimum computer specification.

It is advised to use headphones with working speakers and microphone.

Ajit Jaokar

Ajit is a dedicated leader and teacher in Artificial Intelligence (AI), with a strong background in AI for Cyber-Physical Systems, research, entrepreneurship, and academia. 

Currently, he serves as the Course Director for several AI programmes at the University of Oxford and is a Visiting Fellow in Engineering Sciences at the University of Oxford. His work is rooted in the interdisciplinary aspects of AI, such as AI integration with Digital Twins and Cybersecurity.

His courses have also been delivered at prestigious institutions, including the London School of Economics (LSE), Universidad Politécnica de Madrid (UPM), and as part of The Future Society at the Harvard Kennedy School of Government.

As an Advisory AI Engineer, Ajit specialises in developing innovative, early-stage AI prototypes for complex applications. His work focuses on leveraging interdisciplinary approaches to solve real-world challenges using AI technologies.

Ajit has shared his expertise on technology and AI with several high-profile platforms, including the World Economic Forum, Capitol Hill/White House, and the European Parliament.

Ajit is currently writing a book aimed at teaching AI through mathematical foundations at the high school level.

Ajit resides in London, UK, and he is actively engaged in advancing AI education and innovation both locally and globally. He is neurodiverse – being on the high functioning autism spectrum.

Ajit’s work in teaching, consulting, and entrepreneurship is grounded in methodologies and frameworks he developed through his AI teaching experience. These methodologies help to rapidly develop complex, interdisciplinary AI solutions in a relatively short time. These include:
1. The Jigsaw Methodology for low-code data science to non-developers.
2. The AI Product Manager framework and AI product market fit framework 
3. Software engineering with the LLM stack 
4. Agentic RAG for cyber-physical systems.
5. AI for Engineering sciences: 
6. The ability of AI to reason using large language models

He also consults at senior advisory levels to companies.

His newsletter on AI in Linkedin has a wide following 
https://www.linkedin.com/newsletters/artificial-intelligence-6793973274368856064/

Nicole Königstein

Nicole Königstein is an AI Researcher and Practitioner in Agentic Systems, working across  research, consulting, teaching, and direct system implementation to build reliable, production- ready AI systems. Her work focuses on multi-agent architectures, evaluation, safety, and long-term system behavior.

She served as an external evaluator for a European Commission AI Grand Challenge and has  advised IOSCO on generative AI in regulated environments. She also serves on advisory boards  for leading AI and quantitative finance conferences. Nicole regularly delivers invited talks and technical workshops across academia, industry, and international events.

She is the author of Math for Machine Learning and Transformers in Action with Manning Publications. Her books Transformers: The Definitive Guide: Applications Beyond NLP and AI  Agents: The Definitive Guide have been published by O’Reilly Media, and her forthcoming book Harness Engineering for AI Agents will also be published by O’Reilly Media.

Participants who satisfy the course requirements will receive a University of Oxford digital certificate of completion. To receive a certificate at the end of the course you will need to:

1. Achieve a minimum attendance at online sessions of 75%.
2. Submit completed work as part of the capstone project after the end of the taught sessions

Participants who meet this criteria will be emailed after the end of the course with a link and instructions on how to access their University of Oxford digital certificate. 

The certificate will show your name, the course title, and the dates of the course you attended. You will also be able to download your certificate or share it on social media if you choose to do so.

Fees

Payment

Fees include electronic copies of course materials. 

All courses are VAT exempt.

Register immediately online 

Click the ‘Book now’ button on this webpage. Payment by credit or debit card is required.

Request an invoice

If you require an invoice for your company or organisation, please email us at technology@conted.ox.ac.uk to request an online enrolment form.  Payment is then accepted online, by credit/debit card, or by bank transfer. 

Recommended reading

Koenigstein, Nicole; AI Agents: The Definitive Guide: Design, Deployment, and Evaluation for Production (O’Reilly Media, October 6, 2026) 
https://learning.oreilly.com/library/view/ai-agents-the/0642572247775/ 

Module code: O26C067Q6Y

Please apply by clicking the ‘Book Now’ button below.

If you would like to discuss your application or any part of the application process before applying, please email us at technology@conted.ox.ac.uk.

View our terms and conditions

Book now