Claude Code 2.0 Agentic Coding: No, other agents aren't even close.

On-Device Code Migration with Scout-Plan-Build (2:09 - 3:47)

• The demonstration shifts to migrating a proprietary codebase with eight applications using Claude Code.
• The first step is to activate "thinking mode" for more compute and use a reusable custom slash command /scout plan build.
• The prompt for migration includes documentation URLs and explicit instructions to pay attention to the new syntax for the Claude Agent SDK, particularly the system prompt change.

Understanding the Scout-Plan-Build Workflow (3:47 - 5:04)

• The /scout plan build workflow is a three-step process: scout the codebase, plan the changes, and then build.
• The "scout" step is crucial and involves running custom slash commands within other custom slash commands, allowing for the composition of agentic prompts.
• This composition capability is a key feature of Claude Code 2.0, enabling complex, chained workflows.

Delegating with the R&D Framework (5:04 - 7:19)

• The "scout" prompt's purpose is to delegate the searching process to a fast, token-efficient agent.
• This aligns with the R&D (Reduce and Delegate) framework for context engineering, offloading tasks from the primary agent's context window.
• Four sub-agents (Gemini Light, Codeex, Gemini Flash Preview, Codeex) run in parallel to search for relevant files.
• These sub-agents create a relevant_files.md file, detailing exact file locations and character offsets for updates.
• The scout prompt will be available in the "Building Specialized Agents" codebase for Agentic Horizon members.

Planning and Building Phases (7:19 - 10:14)

• The planner agent reads the relevant_files.md and scrapes necessary documentation locally for all agents.
• The planner's workflow involves analyzing, scraping docs, designing, documenting the plan, generating, saving, and reporting.
• Separating prompts into scout, plan, and build makes workflows clear for humans and agents.
• The ability to chain custom commands in Claude Code is highlighted as an incredible feature for isolating and reusing agentic prompts.
• The build phase executes the plan, migrating the eight custom agents to the new Claude Agent SDK, including critical system prompt migrations and testing.

Context Window Limitations and Solutions (10:14 - 16:05)

• The demonstration reveals that the agentic workflow consumed 51% of the context window, indicating a limitation for single agents.
• The "autocompact buffer" feature in Claude Code 2.0 can consume 22% of the context, but it can be disabled via /config to free up tokens.
• Dedicated AI Developer Workflows (ADWs) are presented as a solution to overcome context window limits by isolating agent context.
• Engineers must understand agent capabilities and limitations (core four: context, model, prompt, tools) to maximize engineering output.

Output Style and Out-of-Loop System Results (16:05 - 21:58)

• IndyDevDan shares a custom output style for Claude Code 2.0 (Observable Tools Diffs TTS) that provides full output, ordered tool calls, and audio task summaries, contrasting with the default truncated output.
• The out-of-loop system on the dedicated agent device successfully completed its task, building three OpenAI Agent SDK prototypes: one with web search, one with a function tool, and one for a real-time agent.
• The web search agent successfully queried for AI developments, and the function tool agent demonstrated local user info retrieval.
• The real-time API agent required further setup for audio implementation, which was handed off for backend work.
• The dedicated agent device provides full logs, mounted GitHub URLs, documentation, and args, allowing for complete traceability of the autonomous engineering work.

The Philosophy of Tactical Agentic Coding (21:58 - 26:29)

• The core idea is to build "the system that builds the system" by investing in reusable agentic prompts and dedicated agent devices.
• This approach offers an "asymmetric return" on engineering time, as complex codebases can be generated with minimal prompts.
• The "Tactical Agentic Coding" course is promoted, covering scaling the core four, elite context engineering, agentic prompt engineering, and building specialized agents.
• The course emphasizes moving beyond simple back-and-forth prompting to building powerful, autonomous systems.

Conclusion (26:29 - 27:53)

• The session, despite running only two main prompts, demonstrated significant agentic capabilities.
• The agent successfully created a 200-500 line proof of concept for the real-time API.
• Viewers are encouraged to focus on important tools, subscribe for future content, and build powerful, autonomous systems.
• Claude 4.5 Sonnet within Claude Code is reiterated as the best agentic coding model for long, end-to-end workflows on local devices and dedicated environments.

"You want to build the system that builds the system." — IndyDevDan