Taming AI Agents With Harness Engineering

// related

• #0001

  #0001Agents & toolsGitHub Changelog24 hours ago

  GitHub adds REST API auditing for `Copilot` cloud agent repo config

  60radar

  Repo-level agent settings can now be checked by API instead of manual UI review. Useful for keeping automation permissions visible before cloud agents touch production code.

  • —New endpoint: Get Copilot cloud agent configuration for a repository, currently in public preview.
  • —Best fit is policy drift checks across repos: scan whether agent access and configuration match your expected defaults.
  • —This is governance plumbing, not a coding-speed feature. Worth adopting if Copilot agents run on real repos.

  Source: github.blog/changelog/2026-05-18-audit-repository-copiloRead original →
• #0002

  #0002Agents & toolsGitHub Changelogyesterday

  `Copilot Spaces API` is now generally available

  70radar

  Copilot SpacesGitHub Copilot feature — manages task-specific context spaces

  Spaces can now be managed from your own apps via API. Useful for wiring repo context into internal tools or repeatable agent workflows.

  • —The API supports create, read, update, and delete for Spaces, so context setup no longer has to stay inside GitHub UI.
  • —Good fit for templates: bootstrap a project space per repo, customer, or feature branch and keep agent context consistent.
  • —This is more automation surface than end-user feature. Value depends on whether Copilot Spaces is already part of the coding workflow.

  Source: github.blog/changelog/2026-05-18-copilot-spaces-api-now-Read original →
• #0003

  #0003Agents & toolsr/ClaudeAIyesterday

  11 Claude Habits That Compound Over Daily Use

  50radar

  The useful part is not prompt tricks, but persistent context: Projects, CLAUDE.md, styles, skills, and subagents. Worth turning into a default setup before long coding sessions.

  • —Put codebase context, style guides, and prior PRs into Projects once. Re-pasting the same background is pure context tax.
  • —A custom style like skeptical senior engineer changes review quality by forcing pushback instead of agreeable code comments.
  • —In Claude Code, CLAUDE.md carries more weight than session prompts. Around 80 lines of project context can remove repeated stack explanations.
  • —Use cheaper/faster models by task: Sonnet as default, Opus for architecture, Haiku for batch cleanup like tickets or emails.
  • —Subagents fit parallel chores: run tests, inspect files, or summarize docs while the main coding thread keeps moving.

  Source: www.reddit.com/r/ClaudeAI/comments/1tgqnsl/11_claude_thiRead original →