Skills, agents and everything else

What was covered

• The problem with coding agents today. Hallucinated tool use, no shared context and results that cannot be reproduced.
• Skills as the building block. Markdown files that define a domain of knowledge, the patterns to follow and the constraints. They give the agent a four-step loop: scan, plan, act, verify.
• A real setup. Felipe shared his own agent configuration: different providers for local and cloud models, with expensive models mapped to complex tasks and cheap ones to simple tasks.
• Orchestrator and sub-agents. An orchestrator delegates to specific sub-agents. Each manages its own context window and returns a summary, which cuts token use and improves accuracy.
• Commands. Custom commands act as shortcuts to complex instructions that use particular skills, which suits recurring jobs such as retrospective reports.

What came up

• Skill or agent? A skill is descriptive knowledge about how to do a specific task. An agent is prescriptive: it defines identity, purpose and the tools it can reach.
• A shared registry. There is a registry of skills for sharing workflows, for example one that evaluates coding challenges and produces PDF reports for reviewers.
• Asking questions. Bruno suggested agents that keep asking until they are fully clear on the task. Felipe prefers models that answer directly over ones that negotiate or stretch the conversation.
• Agents for agile metrics. A team member described using agents with GitHub Copilot and Azure DevOps to manage agile metrics and refine team processes, and the challenge of scaling that to other people.

Takeaway

Share the skills and agent configurations you build in the channel, so the whole team benefits.

This session has no row of its own in the topic sheet, so its date and content come from the meeting notes.