Agentic Programming with Claude Code

Agentic Coding: The Future of Software Development with Agents - YouTube

  • Speaker: Armin Ronacher – creator of Flask & Jinja; long-time open-source maintainer.
  • Focus: “Agentic coding” – working with AI agents that plan, execute and refine code collaboratively, far beyond simple autocomplete.
  • Why the buzz now:
    • Claude 4/Opus models are explicitly trained for tool use.
    • Anthropic’s reference agent Claude Code (terminal-centric) is cheap to try via flat-rate subscription.
    • Dozens of open-source clones sprung up, accelerating experimentation.
  • Key differences from editor helpers like Cursor/Copilot:
    • Agents break big tasks into sub-tasks, run for hours, and let you watch/review steps.
    • Terminal UI + SSH makes it easy to nest agents and script them.
  • Practical lessons from weeks of heavy use:
    • Simpler ecosystems (Go, PHP, “basic” Python) and long, unique function names reduce confusion.
    • Provide fast CLI tools with clear errors; unify logs from server, browser console, SQL, etc.
    • Minimise extra “MCP” plugins; plain command-line tools are easier for agents to compose.
    • Use markdown design docs → agent plans → sub-agents; restart rather than rely on context-compression when context overflows.
    • Add reach()/await markers to coordinate multi-process flows; agents aren’t great at implicit concurrency.
  • CI & dev-ops: Agents can debug GitHub Actions via gh CLI or browser automation; great for “only-fails-in-CI” problems.
  • Everyday automation wins: Downloading videos, placing classified ads, cleaning disk space, tweaking Git config, reading large PDFs via Gemini CLI, nesting agents inside agents.
  • Big picture: The LLM + agent loop feels “catnip for programmers” – highly addictive, still early, and likely to spread far beyond coding. Keep tool subscriptions monthly so you can switch as the landscape shifts.