Claude Code & Agentic Systems — Smitha Shivakumar

Session Overview

What This Session Covers

From agent fundamentals to real multi-agent architectures — with hands-on exercises at every step.

◈

Agentic Foundations

What makes a system agentic — tools, memory, planning, and the difference between a chatbot and an agent that can execute multi-step tasks.

⇄

Claude Code in Practice

Building with Claude Code from the command line — reading codebases, making changes, running tests, and orchestrating tasks end to end.

∿

Multi-Agent Architecture

When one agent isn't enough — designing agent networks, handoffs, shared memory, and parallel execution patterns.

⬡

Safety & Control

Keeping humans in the loop — permission models, confirmation gates, and how to design agentic systems that don't overreach.

Session Structure

The Agenda

Four blocks, two exercises, one mental model shift — from thinking in prompts to thinking in systems.

Block 01

What Is an Agentic System?

The agent loop: perception → reasoning → action → observation — how agents differ from one-shot LLM calls

Tools, memory, and planning — the three primitives every agentic system is built on

The spectrum from assistant to fully autonomous — where your use case falls and why it matters for system design

Block 02

Building with Claude Code

Claude Code setup and the permission model — what it can read, edit, run, and why the controls matter

Task decomposition — how to give Claude Code work it can actually complete autonomously, without micromanaging

Hooks, sub-agents, and parallel execution — the power features most users never reach

Exercise Give Claude Code a real codebase task — scope it, run it, review what it did, and debug where it went wrong

Block 03

Multi-Agent Patterns

Orchestrator + subagent patterns — when to break a task across multiple specialized agents

Shared context and handoffs — how agents pass state without losing information

Parallelization strategies — running independent subtasks concurrently and merging results

Key Principle One orchestrator + N specialists > one generalist agent trying to do everything. Design for separation of concerns.

Block 04

Keeping Control of Autonomous Systems

Risk surface of agentic systems — what can go wrong when an agent has tool access and real-world consequences

Permission tiers and confirmation gates — designing escalation paths for high-stakes, low-reversibility actions

Observability: logging agent decisions, tool calls, and outcomes so you can debug, audit, and improve

Exercise Map the risk surface of an agentic system — classify every action by reversibility, then design the appropriate permission tier for each

What You Leave With

Practical Takeaways

Session Frameworks

Three patterns you can apply immediately

Every framework below is designed to be usable the day after the session — not aspirational, not theoretical.

◈

The Agent Design Checklist

A structured framework for scoping agentic tasks — defining the action space, memory requirements, and human-in-the-loop checkpoints before you write a line of code.

⇄

Multi-Agent Blueprint

A reference architecture for orchestrator + specialist patterns with guidance on handoffs, shared state, and parallel execution strategies.

⬡

The Risk Surface Map

A template for classifying every action an agent can take by reversibility and blast radius — and designing the appropriate permission tier for each.

Skills Covered

What You'll Be Able to Do

Agentic AI Claude Code Multi-Agent Architecture Tool Use & Integration Agent Loop Design Task Decomposition Parallel Execution Human-in-the-Loop Permission Modeling Observability LLM Orchestration

What Is an Agentic System?

Building with Claude Code

Multi-Agent Patterns

Keeping Control of Autonomous Systems

Three patterns you can apply immediately

Bring this session to your engineering team.