Agentic AI: build an agent
Take this if you want to build systems that take actions, and to know what can go wrong before you let one loose.
What you’ll be able to do
You can build an agent with tools, planning and memory, instrument it so its decisions are traceable, evaluate it on a real task suite, and reason concretely about its safety.
What it assumes
The Generative AI phase or equivalent, plus confident Python and experience calling model APIs.
- Price
- ₹1,499
- Length
- 10 weeks
- Sessions
- 10
- Effort
- ~8 h/week · 80 h total
- For
- College · Working professional
Adds it to My courses on your dashboard so you can find it again. Payment is not open yet — nothing is charged.
What’s taught, in order
- 01
The loop, and tool use
An agent as a control loop around a model, then tool schemas, argument validation and the failures that dominate real systems. Written by hand before any framework appears.
- 02
Planning and reasoning
Decomposition, re-planning after failure, and the honest limits of a model's ability to plan — including how to tell a plan from a plausible-sounding list of steps.
- 03
Memory systems
What an agent should remember, where it should live, and how retrieval over its own history differs from retrieval over documents.
- 04
Frameworks and the Model Context Protocol
LangGraph and the wider landscape, then MCP as the emerging standard for connecting tools. Met after you have built the loop yourself, so the abstractions are legible.
- 05
Multi-agent systems
Several agents cooperating, the coordination costs that arrive with them, and the cases where one well-instrumented agent is the better engineering answer.
- 06
Evaluation, observability and safety
Task suites, traces, regression detection, and an agent safety strand covering permissions, sandboxing, prompt injection and the actions an agent should never take unsupervised.
What you build
Build and evaluate an agent that completes a genuinely multi-step task with real side effects, on a suite of at least fifteen task instances. Ship the traces, the pass rate, a categorised failure analysis, and a written safety review naming the actions you gated behind human approval and your reason for each.
Check it yourself against this
- Fifteen or more task instances, with a reported pass rate on all of them.
- Every run produces a trace from which the agent's decisions can be reconstructed.
- Failures are categorised by cause rather than counted.
- At least one guardrail is demonstrated firing on a real attempt.
- The safety review names specific gated actions with specific reasons.
- A prompt injection attempt against your own agent is documented, with the outcome.
Then answer this
Your agent passes most tasks and fails unpredictably on the rest — what would have to change for you to let it run without you watching?
Two or three sentences, in your own words. If you cannot, the course is not finished — go back to the module it came from.
What this course does not do
Single-user agents on tasks you control, not production multi-tenant systems: authentication, rate limiting, cost controls at scale and incident response are out of scope. The safety strand teaches threat modelling and mitigation, not a security certification.
Where this leads
The Agentic AI phase page lists all ten weeks including the MCP and multi-agent weeks, so you can see where framework work sits relative to the hand-built loop that opens the phase.
FAQ
- Which agent framework will I learn?
- LangGraph in depth and several others in survey, but only after week one has you write the loop yourself. That order matters: frameworks make different assumptions about control flow, and you can only evaluate those assumptions if you have made them yourself once.
- Is agent safety a real topic or a disclaimer?
- A real one, with its own week and its own project requirement. An agent with tool access can delete files, spend money and send messages, so permissions, sandboxing and injection resistance are engineering concerns rather than ethics slides.
- Do I need the Generative AI phase first?
- Effectively yes. The memory and retrieval weeks assume you have built a retrieval system and evaluated it, and the planning week assumes you understand sampling well enough to know why a model's plan varies between runs.
Other courses
This is a standalone course. It does not add weeks to the 78-week curriculum or change your roadmap. If you want the whole path instead, the full programme is ₹399/month · ₹2,000/year — see what it covers.