Agentic Coding with Claude Code: Plan Mode, Subagents, and the Cloud Agent

Class Duration

14 hours of live training delivered over 2-3 days to accommodate your scheduling needs.

Student Prerequisites

Professional software development experience in at least one language
Working knowledge of Git and GitHub (issues, pull requests, branches, reviews)
Comfort with the command line and terminal workflows
Anthropic Claude API access or a Claude Code subscription
Familiarity with Claude Code's core features is helpful (see Claude Code in Practice for the feature deep-dive)

Target Audience

Software engineers, tech leads, and platform/DevEx teams who want to move beyond AI autocomplete and delegate engineering work to Claude Code agents. Ideal for teams adopting agentic workflows across the terminal, the IDE, and github.com—assigning issues to Claude in the cloud, driving plan-then-execute sessions locally, and standardizing how agents are supervised, reviewed, and governed. This is a practitioner course focused on agentic development with Claude Code rather than a survey of every feature.

Description

Claude Code is Anthropic's terminal-native coding agent, now reachable from the IDE, the web, and directly inside GitHub. This course teaches developers how to put it to work on real tasks—planning multi-step changes, editing across many files, running tests, and iterating until the job is done. You will learn to launch work directly from github.com by assigning issues to Claude (via Agent HQ) and through @claude mentions powered by the Claude Code GitHub Action, then track and steer the agent as it pushes commits to a draft pull request. We cover plan mode and the permission/auto-mode supervision spectrum, subagents and worktree-isolated parallel work, grounding agents with CLAUDE.md, skills, slash commands, hooks, and MCP servers, remote control that lets a session move from your machine to your phone, and closing the loop with /code-review. Throughout, we emphasize supervision, security, and cost so teams can adopt agentic coding responsibly.

Learning Outcomes

Explain Claude Code's agentic surfaces—terminal, IDE extensions, the web, and the GitHub-hosted agent—and choose the right one per task.
Drive plan-then-execute sessions: plan mode, accept-edits, and classifier-screened auto mode.
Orchestrate subagents and worktree-isolated background work for parallel tasks.
Launch work from github.com by assigning issues to Claude and using @claude mentions.
Understand how the Claude Code GitHub Action runs and produces a draft pull request.
Write agent-ready issues with clear scope and acceptance criteria, and iterate through PR review.
Ground agents in project context with CLAUDE.md, skills, slash commands, hooks, and MCP servers.
Run and coordinate multiple agents, including remote control across machine and phone.
Apply /code-review and route findings back into follow-up fixes.
Establish security, governance, and cost controls for agentic development at team scale.

Training Materials

Comprehensive courseware is distributed online at the start of class. All students receive a downloadable MP4 recording of the training.

Software Requirements

Claude Code CLI installed, an Anthropic API key or Claude Code subscription, a GitHub account with a repository for hands-on labs, the GitHub CLI (gh), a supported IDE with the Claude Code extension (VS Code, JetBrains, or Zed), and Git.

Training Topics

Foundations of Agentic Coding

What "agentic" means: plan, act, observe, iterate
Claude Code's agentic surfaces: terminal, IDE extensions, the web, and the GitHub-hosted agent
The supervision spectrum: permission prompts → accept-edits → auto mode (destructive-action and prompt-injection screening)
Where agents excel (well-tested, low-to-medium-complexity work) and where to keep a human in the loop

Claude Code in the Terminal and IDE

Interactive sessions in the terminal and via IDE extensions (VS Code, JetBrains, Zed)
Multi-step execution: reading, editing, running commands and tests
Reviewing diffs, inline diff comments, and checkpoints/rewind
Steering and interrupting mid-task

Subagents and Background Work

When to spawn a subagent vs. stay in the main session
Subagent scope, isolation, and output collection
Background subagents for long-running tasks
Worktree isolation to run parallel work without conflicts

Planning and Decomposition

Plan mode: agree on an approach before code changes
Turning a goal into agent-sized tasks and todos
Coordinating with a GitHub Spec Kit constitution for spec-driven work
Acceptance criteria the agent can verify against

Grounding Agents in Context

CLAUDE.md for persistent project context and conventions
Skills and custom slash commands for repeatable workflows
Hooks for lifecycle automation and guardrails
MCP servers to connect internal tools, APIs, and data

Launching Work from GitHub and the Web

Assigning issues to Claude on github.com (Agent HQ, public preview)
@claude mentions in issues and PRs via the Claude Code GitHub Action
Claude Code on the web for browser-based async sessions
Starting work from GitHub Mobile and VS Code

How the GitHub-Hosted Agent Works

Running in GitHub Actions via the claude-code-action
Authentication options: Anthropic API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry
Commits pushed to a branch and a draft pull request; session logs
Sandboxed execution, checkpointing, and credential scoping

From Issue to Pull Request

Writing agent-ready issues: scope, context, and constraints
Reviewing the draft PR and reading the agent's reasoning
Iterating with @claude follow-ups and PR comments
Knowing when to take over versus re-prompt

Agentic Code Review

/code-review on the working tree and on pull requests
@claude review in pull requests
Routing review findings into follow-up fix commits
Quality gates before merge

Remote and Multi-Agent Workflows

Remote control: start a session on one machine, continue on your phone
Running multiple subagents in parallel with worktree isolation
Coordinating terminal, web, and GitHub-hosted agents
Headless (non-interactive) execution for scripted automation

Security, Governance, and Cost

Permission modes and allow/deny tool policies
Hooks as safety guardrails (blocking dangerous commands)
Prompt-injection risks in GitHub-triggered runs and mitigations
Secrets and credential scoping in CI
Token usage, cost attribution, and rate-limit management

Workshop

Assign a GitHub issue to Claude and ship the resulting PR to merge
Terminal lab: plan mode → auto mode to build and test a feature
Subagent + worktree parallel task exercise
/code-review → fix-commit exercise
Q&A session