Agentic Coding: The Future of Software Development with Agents - Armin Ronacher

• Create central logging and observability in your dev environment.
• Pull documentation into your repository so you can reference it.

Tools

• Make your tools fast and user-friendly - agents might kill a long-running tool.
• Prioritize logging and observability to give agents context of what is happening: log files, informative error (and success) messages
• Have guardrails in your tools to prevent potential misuse
• Be cautious regarding MCPs; you can often achieve your goal with CLIs (e.g. github) or bash scripts.
  • in fact, it’s even better because you avoid polluting your context window with MCPs and Claude Code can create multi-step scripts using CLIs or other bash scripts.

Context

• Be conservative with context
• A long Claude.md isn’t helpful (will consume to much context)
• Create tools Claude Code can use to gain context and navigate the codebase effectively (e.g. a tool that retrieves all the methods in the Go project, rather than including them in the context from the beginning)
• Provide a way for it to tail the last 20 lines of the central (combined) log
• Consider sub-agents (results aren’t optimal though)
• Results you get after compacting your chat are random (could be good or bad). Might be better to start a new session.

Practical Tips

• Unified log file: combine console.log (browser) + server logs + everything else
• Patch console.log() in the browser -> forward to server via api call.
• Give it a way to log SQL too.
• Provide a make tail-log script to get latest logs.
• Use github cli for powerful stuff: reading PRs, issues, fixing CI (Github Actions), etc.