GitHub Copilot Workspace: GitHub’s Conversational IDE

javi
Pantalla de desarrollo con múltiples ventanas de código en entorno moderno

GitHub Copilot Workspace (announced April 2024, preview later) is GitHub’s attempt to reimagine development workflow: instead of “inline completion”, proposes task-first. Describe what you want, Copilot generates multi-file plan, applies changes, tests. Competes with Cursor Composer but native GitHub. Still in technical preview at time.

The Model

Workflow:

  1. Open GitHub Issue: describes the problem.
  2. Launch Workspace: Copilot reads issue + codebase.
  3. Plan phase: Copilot proposes steps.
  4. Edit phase: Copilot implements changes.
  5. Test phase: runs tests, iterates.
  6. Create PR: with summary.

Essentially AI dev completing tasks end-to-end.

GitHub Integration

Native:

  • Issues as task source.
  • PRs as output.
  • Codespaces as workspace.
  • Actions: CI triggers on changes.
  • Projects: task management.

Enterprise-friendly workflow.

Copilot Workspace vs Cursor Composer

Aspect Copilot Workspace Cursor Composer
Primary interface Browser + GitHub VS Code fork
Task source GitHub Issue Text prompt
Planning phase Explicit Implicit
Multi-file Yes Yes
Ecosystem GitHub-first IDE-first
Maturity Preview GA
Model OpenAI (presumably) Multi-provider

Copilot Workspace: team/GitHub workflow. Cursor Composer: solo developer productivity.

Features

Plan Preview

Before changes, Copilot shows:

  • Affected files.
  • Implementation steps.
  • Dependencies.

User can edit plan before execution.

Multi-File Edits

Single task can touch many files:

  • Add feature in backend + frontend + tests.
  • Refactor across modules.
  • Upgrade library version.

Plan-first approach helps.

Test Integration

After changes:

  • Runs existing tests.
  • Generates new tests.
  • Iterates if fails.

Closed-loop testing.

Codespace Execution

  • Workspace runs in GitHub Codespace.
  • Full dev environment.
  • Standard tools available.
  • Ephemeral.

Typical Session

Issue: "Add user deletion endpoint with soft-delete"

Copilot Workspace:
  Plan:
    1. Add soft_deleted column to User model
    2. Create migration
    3. Add DELETE /users/:id endpoint
    4. Update User serializer
    5. Add tests

  Execute? (user reviews, approves)

Executing...
  - Modified: models/user.py
  - Created: migrations/0005_soft_delete.py
  - Modified: api/users.py
  - Modified: tests/test_users.py

Running tests...
  - All passing.

Creating PR...
  - PR #124: "feat: add soft-delete to user endpoint"

Smooth but quality varies per task.

Where It Shines

  • Well-described issues → clean PRs.
  • Well-tested codebases: Copilot uses tests as spec.
  • Standard tasks: CRUD, bug fixes, simple features.
  • Team workflows: devs review generated PRs.

Where It Fails

  • Ambiguous requirements: needs clearer issue.
  • Complex architecture: misses big picture.
  • Legacy codebases without tests: stumbles.
  • Non-standard patterns: low-quality suggestions.
  • Large refactors: limited context.

Pricing

Availability phases:

  • Technical preview: free limited access.
  • Copilot Enterprise ($39/user/mo): includes workspace features.

Pricing eventually likely matches Copilot Enterprise tier.

vs Cursor for Teams

Cursor:

  • Individual productivity: excellent.
  • Team workflows: limited.

Copilot Workspace:

  • Individual: less streamlined.
  • Team workflows: built-in via GitHub.

For GitHub-centric enterprises, Copilot Workspace makes sense. For individuals, Cursor remains better.

Expected Evolution

GitHub/Microsoft roadmap:

  • Agent mode: autonomous longer-running agents.
  • Custom models: enterprise bring-your-own-LLM.
  • Enhanced testing: more autonomous test generation.
  • Code review: AI reviewer with deeper analysis.

Direction is Copilot → more autonomous.

Security Considerations

  • Code on Microsoft backend: enterprise agreements cover.
  • No training with enterprise code (per terms).
  • IP attribution: clear.
  • Audit logs: enterprise tier available.

Similar concerns to standard Copilot.

Current Limitations

  • Preview: features changing.
  • Quality variance: sometimes perfect, sometimes off.
  • Slow: planning phase adds latency.
  • GitHub-only: no GitLab/Bitbucket equivalents.
  • Primarily English: issues/PRs in other languages limited.

To Start

Requires:

  • GitHub Copilot subscription.
  • Waitlist access for Workspace preview.
  • Repository enrolled in feature.

Request access on GitHub.

Workflow Impact

Team adopting Copilot Workspace:

  • More PRs: AI generates proposals.
  • More review: human validates.
  • Issue quality matters more: vague issues → bad PRs.
  • Critical test coverage: tests as spec.

Changes incentives in useful direction.

Comparison with AI Agents

Copilot Workspace is specific “agent”:

  • Goal: complete task from issue.
  • Tools: code edit, test, commit.
  • Feedback loop: test → iterate.
  • Scope: typically single PR.

For multi-repo or cross-project, other approaches.

Conclusion

Copilot Workspace is GitHub’s interesting bet on task-first AI development. For GitHub-centric teams, natural fit. For individual productivity, Cursor probably more refined. Preview state means expected evolution. Industry direction — AI doing more autonomously vs just completing inline — is clear. Workspace positions GitHub/Microsoft for that future. Worth evaluating if you have Copilot Enterprise or are in GitHub ecosystem.

Follow us on jacar.es for more on AI coding, GitHub, and development workflows.

Entradas relacionadas