The practical AI dev workflow: which tool for which job

Three architectures, not three flavors of the same tool. If you're comfortable in a terminal, want CI integration, and think in diffs, Claude Code rewards that fluency directly. If you live in VS Code and don't want to leave the editor, Cursor has the smallest learning curve because the shell is one you already know. If you're managing several agents on parallel tasks and want a fleet dashboard rather than one agent in one window, that's Devin Desktop — the product formerly known as Windsurf, which is a rename most 2025-era comparisons still get wrong.

If GitHub Copilot is already in your stack, check your tier before buying a second agent: Copilot's Pro+ ($39/month) and Max ($100/month) tiers both bundle third-party agent access that GitHub names explicitly as "Claude Code, Codex." The full pricing table and where the two tools' architectures actually diverge live in Claude Code vs GitHub Copilot.

Whichever agent you pick, the gap between a plausible-looking diff and one that's actually correct comes down to whether it can check its own work before telling you it's done. Getting real work out of AI agents covers the concrete, currently-documented pieces: Stop hooks that mechanically block a turn until a test/build/lint check passes, subagents that keep noisy exploration out of your main context, and worktree isolation so two parallel sessions can't silently stomp on each other's edits. None of this makes an agent infallible — it makes its failure modes visible instead of silent.

Model Context Protocol is how an agent reaches your actual tools and data instead of guessing from training data alone. The catch: a lot of "here are the official MCP servers" posts point at a repo layout that's no longer current — GitHub, PostgreSQL, and Slack have all moved to an archived repo, and the currently-maintained reference servers are a much shorter list. MCP servers explained has the up-to-date list plus the actual security question worth asking before you approve any server's install command: a local MCP server runs with your full privileges, not sandboxed.

An agent that reads your codebase is only half the workflow — the other half is where your own notes, decisions, and research actually live. Notion vs Obsidian vs Logseq breaks down the real fork in the road: a proprietary cloud database with strong native collaboration (Notion), or plain Markdown/Org files on your own disk that are git-able the same way your code is (Obsidian, Logseq). If an AI layer specifically over your notes matters to you, AI note-taking tools worth paying for prices the options that do one checkable job — transcription, source-grounded Q&A — over the ones making a vaguer "organizes your thinking" promise you can't audit.

If a developer-oriented, git-able vault template appeals more than starting from a blank Obsidian install, we also operate Vaultsmith store — a disclosed house link, not a paid placement; see the footer disclosure.