Copilot is just a fancy autocomplete (vol.3) - Steps towards agentic development

Key points

• Move from general-purpose agents to custom agents with scoped tools and personas for specific workflow tasks like planning, coding, testing, or documentation.
• Use skills to avoid agents rediscovering the same procedures repeatedly; skills load progressively to avoid bloating the context window unnecessarily.
• Implement reflection loops by asking agents to revise their instructions and skills based on mistakes they make, so they improve automatically over time.
• Employ sub-agents with isolated context windows to parallelize research tasks, returning condensed summaries to the main agent instead of raw transcripts.
• Run background agents via Copilot CLI in worktree isolation mode for long-running tasks, and integrate agents into terminal scripts and pipelines for true workflow automation.

Quotes

• “Fix the agent. Don’t fix the code.” (Summarizes the core improvement strategy: agents learn and adapt when instructed on mistakes rather than developers manually fixing the output each time.)
• “Claude Sonnet and Opus, they’re the best for AL. Though, I’ve seen Gemini 3.1 Pro come very close.” (Identifies which models perform best for Business Central AL development at the time of writing.)

AL objects mentioned

As named in the post. A name that matches one object page by exact type and name links to it; the others stay as named.

• other “AL Symbols MCP”
• other “BC Container Helper”
• other “TelemetryBuddy”
• other “BC Code Intelligence”

Context

• Features: Custom agents, Agent skills, Reflection loop, Sub-agents, Background agents, GitHub Copilot /create-agent, Copilot CLI, Worktree isolation

Source: Tech Adventures in Business Central, community blog. Summary, key points and quotes are derived (CONTENT-NOTICE.md); read the original for the full text.