Overview
GitHub Copilot spans inline completions, IDE chat agents and a cloud coding agent that works on issues and opens pull requests. The productive use is delegation — clear scope, repository context, human review — and the monthly AI credit system meters agent work while completions stay free on paid plans.
From autocomplete to agent platform
Most developers still use GitHub Copilot the way it launched: accept the grey suggestion, tab, move on. The product has moved past that. Copilot is now a stack: inline completions, chat agents inside the IDE, a cloud coding agent that works on GitHub itself, code review on pull requests, and a terminal CLI — with model access beyond OpenAI through its own routing.
The difference between the old and new use is not a feature list, it is a workflow. Autocomplete waits for you to type and suggests the next line; agent mode takes a task, plans the steps, edits across files, runs the build and fixes its own errors. The second is a different kind of tool with a different skill requirement — and most users never make the switch, which is why the tool's ceiling is higher than most people's usage.
The mindset shift: delegate, don't type
The framing that changes results: treat Copilot like a fast junior engineer, not a keyboard shortcut. Junior engineers do great work with clear scope, repository context and a definition of done — and produce garbage when told "fix it" and walked away from. The leverage is in writing better instructions, not accepting more suggestions.
The mechanism that makes that concrete is the custom-instructions file. Add a file at the repo root and every chat and agent request starts with that context:
# .github/copilot-instructions.md
Python 3.12 FastAPI service. Use pytest, not unittest.
All new endpoints need a rate-limit decorator.
Never log request bodies.
Ten minutes of writing, permanent payoff: every generation in that repo now knows the stack, the conventions and the things the model keeps getting wrong. The next step is delegation at GitHub scale — assign an issue to the cloud coding agent the way you would assign it to a person, and it researches the repo, makes changes on its own branch, runs security scanning and opens a draft pull request. Teams report the same pattern: assign three small issues before a meeting, come back to three PRs. The human job narrows to scope at the start and review at the end — which is the honest description of where AI coding is today, and the review step is not optional, because agents produce confidently wrong logic just like junior engineers do.
What the credit system changed
Pricing moved to a usage-based model: every paid plan includes a monthly allotment of AI credits, while code completions are free on all paid plans. The plans: Free (2,000 completions a month plus limited chat), Pro at $10/month, Pro+ at $39, and Max at $100 with a larger credit pool; teams pay per user on Business and Enterprise.
The practical effect: completions are effectively unlimited for most users, and the credits are the meter on the expensive part — agent runs, chat with premium models, cloud agent work. Heavy agent users plan around the pool; completion-only users never notice it. The credit system is the clearest signal of where the product's real cost sits: the marginal cost of suggestions is near zero, and the cost of autonomous work is real enough to meter.
The delegation checklist before assigning work to the agent:
- An instruction file defines the stack and conventions
- The task has a definition of done
- Scope is bounded to something reviewable in one sitting
- The draft PR will get a human review pass
Where it fits
- ✓ Works for: developers who will invest in instructions and review discipline — that is where the time savings are; teams using issue-driven workflows, where the cloud agent converts small bug fixes into draft PRs; organizations already on GitHub, where Copilot sits inside the existing workflow instead of adding a new one.
- ✗ Not a fit for: developers who want zero-config AI — the tool rewards the setup effort, and users who skip it get autocomplete while paying for the platform; anyone expecting reliable output without review, since agent work still needs human verification like a junior's PR would; non-GitHub-centric teams, where the platform advantages (issue integration, code review) are the weakest.