Agentic Coding

Claude Code 101 is a beginner-level course from Anthropic Academy that covers advanced context management and agentic workflows for developers. The course teaches memory management through CLAUDE.md files for persistent project memory, subagent creation to delegate tasks and maintain clean context, and context window optimization using /compact, /clear, and /context commands. It also covers the explore → plan → code → commit workflow, custom tool integration via MCP servers, and hooks for deterministic control. The course includes modules on prompt engineering patterns, code review practices, and authorization frameworks relevant to building AI coding agents.

Anthropic's Thariq Shihipar discusses advanced prompting techniques including SCQA prompting and decision notes for steering frontier AI agents, noting that Claude often considers correct solutions in reasoning but discards them before execution. Claude Mods were built to run directly inside Claude Code's TypeScript runtime using in-process TypeScript and cached forked agents for mutable software development. Context files like CLAUDE.md and AGENTS.md are predicted to become obsolete as frontier models improve due to Pareto dominance in reasoning models and effort allocation dynamics that overconstrain agents.

Claude Sonnet 5.5 costs half the list price of Opus 5.5 ($2/$10 vs $4/$20 per million tokens) and outperforms it on Terminal-Bench 4.0 (70.6% vs 66.4%), an agentic coding benchmark. On most other benchmarks, Sonnet 5.5 lands within two points of Opus 5.5. The actual per-turn cost gap is closer to 1.65× rather than 2× when accounting for identical cache read costs. Sonnet 5.5 is recommended as the default for well-scoped agentic coding tasks, bug fixes, and terminal work at medium effort, while Opus 5.5 remains stronger for complex, open-ended work requiring sustained judgment. A critical architectural consideration: Sonnet 5.5 and Opus 5.5 cannot read each other's thinking blocks, so routers that switch between them mid-conversation silently drop the previous model's reasoning. The recommended approach is to choose one model per conversation, escalate to a harder model by starting a fresh conversation from a summary, or use the advisor tool to consult Opus without switching. Five breaking changes occur when upgrading from Sonnet 5, including disabled thinking type (replaced with between_tools), forced tool_choice restrictions, updated tool IDs, advisor pairings, and append-only conversation history enforcement on new accounts.