GAI201INTERMEDIATE
Applied Agentic AI Systems
Orchestrate agents. Prove they work. Run your role on them
A useful agent needs more than a clever prompt. Learn to connect steps, set boundaries and test what happens when things do not go to plan.
- Audience
- Experienced GenAI users
- Duration
- 18.5 hours
- Format
- 3 days · In person
- Level
- Intermediate
- Coding
- No coding required
- Prerequisite
- GAI101, GAI101F or equivalent working generative-AI fluency
- Full fee
- S$1,750
01What you’ll build
No coding required · 3 days · In person
What you’ll build: A role system map, three running multi-agent patterns, an agentic workflow graph and evaluation and operations evidence.
A one-page role system map
Map three AI components, shared memory, routines, contracts and autonomy boundaries for your role.
Three running agent patterns
Build a pipeline, parallel fan-out and orchestrator–workers system on real work inputs.
Evaluation and operations evidence
Build a set of test examples to check your workflow, then ship a regression result, permissions, approval gate, runbook and autonomy decision.
02Curriculum
18.5 hours
Curriculum
Day 01
- Map recurring responsibilities across memory, routines, reactive agents, an on-demand harness and governance.
- Choose three components and give each a contract covering its goal, inputs, output, constraints and definition of done.
- Build one component twice: first as a single agent, then as an orchestrator with two specialist agents.
- Compare quality, time and token use to decide whether the multi-agent design earns its complexity.
- Build a pipeline with a quality gate between stages.
- Build a parallel fan-out with independent workers and an aggregation step.
- Build an orchestrator–workers system in which the lead decomposes and delegates the task.
- Define tool permissions and identify when each pattern is the wrong shape for the work.
Day 02
- Draw an agentic workflow as nodes and edges, including branches, human review and failure paths.
- Wire the graph in n8n with routing, a parallel join, a bounded loop, an approval gate, an error edge and a sub-workflow.
- Design durable state and per-node context so runs can be inspected, interrupted and resumed.
- Represent the same workflow as a script and expose the n8n workflow through an MCP bridge.
- Diagnose seeded teams that loop, stall or produce confidently wrong output.
- Trace failures to missing context, weak contracts, wrong tools, absent stopping conditions or poor verification.
- Repair each failure structurally with clearer authority, bounded retries, durable state, permissions or mechanical checks.
- Re-run and re-measure the systems, then encode each useful lesson in both guidance and a check.
Day 03
- Build a 20-case golden set from realistic work inputs and define a scorable rubric.
- Compare human scores with an LLM-as-judge run and inspect where the judge disagrees or shows bias.
- Read traces, tool calls, cost and change receipts to evaluate both the output and the path taken.
- Change one model, prompt or routing rule, run a regression and make an evidence-backed ship or no-ship decision.
- Set a token budget, retry limit, kill switch and least-privilege tool permissions.
- Place a human approval gate before the sensitive action and test it with a poisoned-document injection drill.
- Assign autonomy levels using the evaluation evidence, keeping autonomy within the strength of the harness.
- Write a one-page runbook and complete a peer review against the 12-point reliability checklist.
03Your instructor
Metacamp trainer
Your instructor: Vincent Leow
Tech Lead, Neutral Trade.
As a current tech lead and former DeFi protocol co-founder, he brings hands-on experience architecting, testing and operating production software systems.
04Course FAQ
A few things worth knowing before you choose.
Still unsure? Write to team@metacamp.sgFull fee
S$1,750
Funding to be confirmed
Start with GAI101 if you are new to applied generative AI. Choose GAI101F to build within a financial-services context, AI101 to analyse financial data with Python, or GAI201 after GAI101 or equivalent experience to orchestrate multi-agent systems.
GAI101 and GAI101F require no prior AI or coding experience. AI101 teaches Python from the fundamentals and requires tertiary education. GAI201 requires GAI101, GAI101F or equivalent working generative-AI fluency, but no coding experience.
Funding is course-specific. Eligible Singapore Citizens may use up to S$500 from their Opening SkillsFuture Credit toward the remaining fee. SkillsFuture Mid-Career Credit cannot be used. Available subsidies and final fees depend on the approved course run and your eligibility; confirm the prevailing amount before enrolment.
Something you built and know how to use. Depending on the course, this may be a custom assistant and automation, a financial-services AI workflow, a Python data-analysis capstone or a tested multi-agent system.
Yes. We offer private programmes for organisations and functional teams. Examples and exercises can be shaped around relevant roles, workflows and business goals.
AI will continue changing after your course. We teach principles and ways of thinking that remain useful as tools evolve. Graduates can also continue learning with the Metacamp community as new sessions and programmes are announced.