Home › Tracks › Agentic Systems Engineering
12-week elite track
Orchestrate fleets of autonomous agents that ship real work.
Engineer reliable multi-agent systems: tool use, planning, memory, sandboxed code execution, and orchestration patterns that move from a single augmented LLM to a coordinated fleet. You build agents an enterprise can put in production, not demos. Bridges directly to classic Operating Systems, Distributed Systems, and Compilers.
Week by week
Every week unlocks the next. Concepts route you to free, world-class material; projects turn that knowledge into something deployed.
Week 1
An agent is an LLM in a loop with tools, retrieval, and memory. Master the atomic building block before composing it: when a deterministic workflow beats an agent, and the failure modes of handing a model autonomy.
Bridges to Operating Systems — processes, scheduling, and the run loopBuilds on: nothing — start here
Week 2
Give a model hands: structured tool schemas, argument validation, error feedback loops, and parallel tool calls. The contract between a model and the outside world is an API design problem.
Bridges to Software Engineering — interface design and API contractsBuilds on: The Augmented LLM as a Building Block
Week 3
Standardize how agents connect to tools, data, and other agents. MCP servers and clients, transports, capability negotiation, and why a protocol beats bespoke integrations at scale.
Bridges to Computer Networks — protocols, client-server architecture, and RPCBuilds on: Tool Use & Function Calling
Week 4
How agents break a goal into steps: chain-of-thought, ReAct, plan-and-execute, and reflection. When explicit planning helps and when it just adds latency and drift.
Bridges to Artificial Intelligence — search, planning, and state-space reasoningBuilds on: Tool Use & Function Calling
Week 5
Agents forget. Engineer short-term scratchpads, long-term vector memory, summarization, and context-window budgeting so an agent stays coherent across a long-running task.
Bridges to Operating Systems — memory hierarchy, paging, and cachingBuilds on: Planning, Reasoning & Task Decomposition
Week 6
Letting an agent run code it wrote is a security problem first. Isolation via containers and microVMs, filesystem and network confinement, resource limits, and safe failure.
Bridges to Operating Systems — virtualization, namespaces, and process isolationBuilds on: Agent Memory & Context Management
Week 7
You are handed an enterprise mandate: the research division needs a launched product — a system that takes an open-ended technical question, autonomously researches it across many sources, synthesizes a defensible report, and ships the report as a published artifact, with zero human steps in the middle. Build an orchestrator-worker multi-agent system: a lead agent that decomposes the question and spawns specialized worker agents (search, read, synthesize, fact-check), coordinates their results through shared state, and produces a cited deliverable. This is not a notebook demo. The result must be a directly deployable, hyperscalable product: real public hosting, CI/CD on every commit, observability, security hardening, a polished and accessible web UI a non-technical analyst will happily use, and complete go-to-market material — a landing page, a pitch, and a recorded demo. The architecture must absorb concurrent research runs without falling over, and recover from a failed worker. We are not here to babysit the run; ship it as a real product.
Why it matters: Multi-agent research and synthesis systems are being deployed across consulting, finance, and R&D to compress weeks of analyst work into hours. Shipping a coordinated, fault-tolerant agent fleet makes a builder ready for an Agentic Systems Engineer or Applied AI Engineer role at the ₹1-crore tier, where the bar is production reliability, not a demo.
A publicly hosted product with its own domain or stable URL, plus a public repo: the orchestrator and worker agents, an MCP-based tool layer, a fast accessible web UI for submitting questions and reading results, CI/CD running lint/tests/build on every commit, persisted and inspectable run traces, a marketing landing page, a 10-slide pitch, a recorded demo video, and a README documenting the coordination design, the failure-recovery and scaling strategy, and three example end-to-end runs with their published reports.
Market signal — who wants thisAgentic deep-research is one of the hottest 2026 categories: the AI agent market is projected to grow from $7.84B in 2025 to $52.62B by 2030 (41% CAGR), and a16z reports a portfolio pivot from copilots to autonomous systems, with Sierra, Glean, and Decagon as comparables and YC W26 funding multi-agent orchestration startups such as Tensol and Korso. Consulting, finance, and corporate R&D teams are actively buying systems that compress weeks of analyst work into hours; investors fund this because it sells time back to high-cost knowledge workers.
Week 8
Compose specialized agents into a system: orchestrator-worker, handoffs, shared state, and message passing. The coordination problems are the same ones distributed systems have always had.
Bridges to Distributed Systems — coordination, message passing, and consensusBuilds on: Sandboxed Code Execution
Week 10
Agents fail silently and non-deterministically. Build trajectory evals, tool-call assertions, LLM-as-judge, and tracing so you can tell a regression from noise.
Bridges to Software Engineering — testing, regression suites, and observabilityBuilds on: Multi-Agent Orchestration
Week 11
An autonomous agent with tools is an attack surface. Prompt injection, tool poisoning, the lethal trifecta of private data plus untrusted content plus exfiltration, and least-privilege agent design.
Bridges to Information Security — threat modeling and least privilegeBuilds on: Evaluating & Observing Agents
Week 12
An enterprise mandate: ship a launched product — an autonomous coding agent that, given a real GitHub issue, plans a fix, writes and runs code inside a hardened sandbox, iterates against test feedback, and opens a pull request, with untrusted code never touching the host. This is an operating-systems problem as much as an AI problem: the agent's generated code is hostile by assumption. Confine it with containers or microVMs, enforce filesystem and network policy, cap CPU and memory, and survive an agent that tries to escape or hang. The deliverable is not a script — it is a directly deployable, hyperscalable product: real public hosting, CI/CD, observability, a clean dashboard where a developer queues issues and watches trajectories, security hardening, and full marketing (landing page, pitch, demo). The sandbox fleet must scale horizontally to many concurrent jobs. We are not here to babysit a run; ship it as a real product.
Why it matters: Autonomous coding agents are now a core part of engineering org tooling, and the hard part is safe execution at scale, not code generation. A builder who can ship a sandboxed, evaluated coding agent is directly deployable as an Agent Infrastructure Engineer or AI Platform Engineer, roles commanding ₹1-crore-plus compensation because they sit between security and AI.
A publicly hosted product with a stable URL, plus a public repo: the planning-and-execution loop, the sandbox isolation layer, a fast accessible dashboard for queuing issues and inspecting trajectories, CI/CD on every commit, production observability, an eval suite over a set of real issues with pass/fail trajectories, a marketing landing page, a 10-slide pitch, a recorded demo, and a README documenting the isolation threat model, the horizontal-scaling design, and the workflow-versus-agent decision.
Market signal — who wants thisAutonomous coding agents are the breakout 2026 developer-tools category: Cursor reached $2B ARR in February 2026 and is raising at a $50B+ valuation, and Replit raised a $400M Series D at a $9B valuation in March 2026. SWE-bench Verified is now the industry yardstick, with top agents exceeding 80%. A whole sub-market of sandbox-execution infrastructure (E2B, Northflank, Modal, Blaxel) is being funded specifically to run agent-written code safely; investors back this because safe execution at scale, not code generation, is the unsolved bottleneck.