The tool market wants you subscribed to five assistants. Your workflow only has three jobs for them. Most “my stack is a mess” problems are not skill problems - they are overlap problems, and they cost you money and attention every day.
The three jobs
Every coding agent you can buy today is doing one of these:
- The interactive loop. You steer each change: completions, inline edits, diff review. The tool lives where you edit.
- The delegation loop. You hand over a whole task - brief in, verified result out. The tool lives where your code runs: a terminal, a sandbox, a CI job.
- The review layer. Suggestions arrive where work gets merged: pull requests, issues, code review comments.
A solo builder can cover all three with two tools: one editor-native assistant for the interactive loop, one terminal agent for delegation, with review folded into whichever one touches your PRs. Teams add a dedicated review layer because more people need the suggestions where they already look.
Three starter stacks
- Starting out (0 $/month to learn): one freemium editor assistant. Learn to spec, review diffs, iterate. Adding a second tool now only adds confusion.
- Shipping real projects: editor assistant + one terminal agent. The assistant keeps your daily flow fast; the agent takes the tasks you can describe end to end. They share the same repo and the same
AGENTS.md, so nothing is locked in. - Team on GitHub flows: editor assistant for individuals + review-layer agents on PRs. Delegation comes later, once your test suite can vouch for merged work.
Decision rules before you add a tool
Run every candidate through these four questions. If it fails one, don’t add it.
- Which job does it do? If the answer is “the same one my current tool does, but smarter”, you are shopping, not solving.
- Can it read my conventions? No
AGENTS.mdor rules-file support means you re-teach it your repo every session. - Can it verify its own work? A delegation tool that cannot run your tests is a suggestion machine.
- What does leaving cost? Prefer tools that work on plain repos and standard files. Your escape hatch is your leverage.
The two traps
- Paying for overlap. Two chat assistants that both autocomplete the same file deliver zero combined value minus one subscription.
- Tool-switching as procrastination. If output quality is the problem, the fix is almost always the brief, the context file or the test loop - not the logo. Switch tools when a job is uncovered, never when a result disappoints.
The takeaway
Decide the three jobs. Buy one tool per job. Put the money you save into the parts that transfer between every tool you will ever use: your spec habits, your AGENTS.md, and a test suite an agent can run.