.webp)
.svg)
Program overview
Online
On-site
Hybrid
Agentic AI Bootcamp: Building Production Multi-Agent Systems
A five-day intensive on agentic systems that survive production. Covers agent design patterns, tool and memory architecture, MCP and A2A protocols, orchestration topologies, agent evaluation, guardrails and human oversight, agent security and enterprise governance. Assumes prior LLM application experience or completion of the AI Engineering bootcamp.
Duration:
5 Days
Level:
Advanced
.avif)
1500+ users onboarded
Who will Benefit from this Training?
Training Objectives
Build a high-performing, job-ready tech team.
Personalise your team’s upskilling roadmap and design a befitting, hands-on training program with Uptut
Key training modules
Comprehensive, hands-on modules designed to take you from basics to advanced concepts
Download Curriculum
- What Makes a System Agentic
- The observe, reason, act and reflect loop and how it differs from a workflow
- Why agent loops broke before 2026 and what changed to make them hold
- The autonomy spectrum from suggestion through to unsupervised execution
- Choosing the lowest autonomy that solves the problem
- The cases where an agent is strictly worse than a script
- Agent Design Patterns
- Reason-and-act loops and plan-then-execute designs
- Reflection and self-critique patterns, and their real limits
- Supervisor and worker delegation
- Evaluator and optimiser pairs for quality-sensitive work
- Routing, map-reduce and decomposition patterns
- Hands-on: implement the same task with two patterns and compare
- Tool Design for Agents
- Designing tool schemas an agent invokes correctly first time
- Error surfaces: what a tool should return when it fails
- Idempotency, retries and side-effect safety
- Sandboxed code execution and isolation for untrusted output
- Hands-on: harden a badly designed tool until an agent uses it reliably
- Context Engineering for Long-Running Agents
- Context window management as the dominant quality constraint
- Compaction, summarisation and lossy compression of history
- Subagents as context isolation rather than just parallelism
- Deciding what an agent needs to remember within a single run
- Hands-on: extend an agent's effective working session without quality collapse
- Memory Architecture
- Working, episodic and semantic memory as distinct concerns
- Why agent memory is transient and what that forces you to build
- The shared context layer as persistent state across pipeline steps
- Retrieval into memory, and memory as a tool the agent calls
- Staleness, contradiction and pruning strategies
- Hands-on: add a persistent memory layer to a multi-step agent
- Agentic Retrieval
- Retrieval as a tool the agent decides to use rather than a fixed pipeline stage
- Query planning, decomposition and iterative retrieval
- Knowing when to stop searching and answer
- Grounding and citation in an agentic flow
- Hands-on: convert a static RAG pipeline into an agentic one and compare
- MCP: The Tool and Data Layer
- MCP primitives: tools, resources and prompts
- Building a server that exposes internal capability cleanly
- Transports, clients and deployment options
- Registry, discovery and governing which servers are approved
- Hands-on: build and register an MCP server against an internal API
- A2A and Agent Interoperability
- How agent-to-agent communication differs from tool access
- Agent discovery, capability advertisement and task delegation
- Transport, authentication and role-based access in cross-agent calls
- The connectivity explosion problem as agent counts grow
- Interface protocols for streaming agent state to a user interface
- Hands-on: expose an agent over a delegation protocol and call it from another
- The Framework Landscape
- Graph-based orchestration and why an explicit state machine prevents runaway loops
- Role-based multi-agent frameworks and where they fit
- Vendor agent SDKs and the lock-in trade-off
- Building without a framework, and when that is the right call
- Hands-on: implement one system in two frameworks and assess the difference
- Orchestration Topologies
- Supervisor and worker: central routing, easiest debugging, bottleneck risk
- Peer-to-peer coordination and shared protocols of conduct
- Hierarchical orchestration for deep task trees
- The centralised control plane as the dominant enterprise pattern
- Single points of failure and multi-region considerations
- Choosing a topology from task structure rather than preference
- Coordination Failure Modes
- Context inconsistency as the primary cause of production failure
- Deciding which agent holds authority over shared definitions
- Conflict resolution between specialist agents
- Loop detection, maximum depth, token budgets and timeouts
- Cost explosions and how they usually happen
- Hands-on: break a working multi-agent system deliberately, then fix it
- Agent Evaluation
- Evaluating the trajectory, not only the final answer
- Rubric grading for multi-step reasoning and tool selection
- Simulation and scripted environments for repeatable testing
- Regression suites for agent and prompt changes
- Measuring cost and latency per completed task as quality dimensions
- Hands-on: build a trajectory evaluation suite
- Guardrails and Human Oversight
- Human-in-the-loop against human-on-the-loop supervision models
- Approval gates and where in a flow to place them
- Blast-radius containment and reversible-by-default design
- Interrupt, cancel and kill-switch mechanics for long-running agents
- Designing for the case where the agent is confidently wrong
- Agent Security
- Direct and indirect prompt injection through tools, documents and web content
- Tool poisoning and untrusted connector risk
- Excessive permissions and privilege escalation paths
- Data exfiltration routes and how to close them
- Threat modelling an agent before it ships
- Hands-on: attack a running agent, then harden it
- Governance, Identity and Observability at Scale
- Agent identity and applying existing access policy to non-human actors
- Inventory and registry: knowing what agents exist and who owns them
- Delegation tracing: who did what, at what cost, through which chain
- Audit trails and decision traceability for regulated environments
- Ownership, review cadence and safely retiring an agent
- Capstone Build
- Design and build a multi-agent system against a real business process
- Wire tools through MCP and delegation through an agent protocol
- Attach memory, trajectory evaluation and guardrails
- Threat model the system and demonstrate the mitigations
- Present the architecture, failure analysis and governance model
Hands-on Experience with Tools
Training Delivery Format
Flexible, comprehensive training designed to fit your schedule and learning preferences
Opt-in Certifications
AWS, Scrum.org, DASA & more
.avif)
100% Live
on-site/online training
.avif)
Hands-on
Labs and capstone projects
.avif)
Lifetime Access
to training material and sessions
Skill-Gap Assessment
Analysing skill gap and assessing business requirements to craft a unique program
1
Personalisation
Customising curriculum and projects to prepare your team for challenges within your industry
2
Implementation
Supplementing training with consulting support to ensure implementation in real projects
3
Why this course
- Built on production failure modes: Structured around what actually breaks at scale, not around what demos well.
- Protocols as first-class content: MCP and A2A taught properly as the two-layer backbone of interoperable agent systems.
- Memory gets its own day: Context inconsistency is the main cause of multi-agent failure, so it is treated seriously.
- Governance included, not bolted on: Identity, audit and oversight designed in, because that is what blocks enterprise rollout.
Training objectives
- Judge when a task genuinely warrants an agent and when it does not
- Apply the core agentic design patterns and select between them deliberately
- Design tools and sandboxed execution that agents can use reliably
- Engineer context for long-running agents through compaction and delegation
- Architect working, episodic and shared memory with a persistent state layer
- Build MCP servers exposing internal tools, data and prompts
- Use A2A and related protocols for cross-vendor agent interoperability
- Compare LangGraph, CrewAI and vendor agent SDKs and choose on real criteria
- Design supervisor, peer-to-peer and hierarchical orchestration topologies
- Diagnose and prevent coordination failures, loops and runaway cost
- Evaluate agents on trajectory as well as final output
- Implement guardrails, oversight models and blast-radius containment
- Defend against prompt injection, tool poisoning and privilege escalation
- Establish agent identity, inventory, audit and cost accountability at scale
Who will benefit
- AI and machine learning engineers building agent systems
- Backend engineers and solution architects
- Platform and security engineers supporting agent deployment
- Technical leads accountable for agents reaching production
.avif)
Lead the Digital Landscape with Cutting-Edge Tech and In-House " Techsperts "
Discover the power of digital transformation with train-to-deliver programs from Uptut's experts. Backed by 70,000+ professionals across the world's leading tech innovators.
.svg)
.svg)
.svg)
.svg)
.svg)
Recommended Training
Continue your learning journey with these hand-picked courses and bundles
Frequently Asked Questions
2. Will my team get any practical experience with this training?
.svg)
With our focus on experiential learning, we have made the training as hands-on as possible with assignments, quizzes and capstone projects, and a lab where trainees will learn by doing tasks live.
3. What is your mode of delivery - online or on-site?
We conduct both online and on-site training sessions. You can choose any according to the convenience of your team.
4. Will trainees get certified?
Yes, all trainees will get certificates issued by Uptut under the guidance of industry experts.
5. What do we do if we need further support after the training?
We have an incredible team of mentors that are available for consultations in case your team needs further assistance. Our experienced team of mentors is ready to guide your team and resolve their queries to utilize the training in the best possible way. Just book a consultation to get support.