GitHub Copilot¶
Tools · 30 min · no code
Before this: Developer tools · After this: Claude Code
Abstract
Copilot spans inline completion, a chat pane, an agent mode that edits across files, a standalone terminal CLI, and a coding agent that opens pull requests. This page covers what each is for, and is explicit about the two products that were retired — because a lot of published material still teaches them.
Verified as of 2026-08-21.
Two things this page used to cover no longer exist
gh copilot suggest / gh copilot explain — the GitHub CLI extension is
retired. GitHub's own documentation says so plainly. It was replaced by a
standalone GitHub Copilot CLI (the copilot command).
GitHub App-based Copilot Extensions — the @mention extension platform
was deprecated in September 2025 and shut down on 10 November 2025,
replaced by MCP servers.
If you find a tutorial building a Copilot Extension as a GitHub App with webhooks and SSE streaming, it is describing a platform that no longer runs. Extensibility is now MCP — see Model Context Protocol.
The surfaces¶
| Surface | What it does | When to reach for it |
|---|---|---|
| Inline completion | Completes the line or block you are typing | Constant, ambient. Boilerplate, obvious next lines |
| Copilot Chat | Conversational, workspace-aware, in a side pane | Explaining unfamiliar code, generating a scaffold |
| Agent mode | Multi-file edits, runs commands, iterates | A change that spans files |
| Copilot CLI | The copilot command in a terminal |
Shell tasks, away from an editor |
| Coding agent | Runs on GitHub, opens exactly one PR per task | Well-specified work you can review async |
| Code review | Reviews a PR diff | A second pass before a human review |
Inline completion is the one people mean when they say "Copilot", and it is the least interesting. Agent mode and the coding agent are where the behaviour is genuinely different — they run a loop, which is the subject of the agent loop.
Getting good results from completion¶
Open the files that matter. Completion draws on your open tabs. If you want it to follow a pattern, have the file containing that pattern open.
Write the signature and the docstring first. A named function with a stated contract produces far better completions than an empty body.
Name things precisely. calculateTaxForInvoice gets you a better completion
than calc. The name is most of the prompt.
Reject fast. Reading a wrong suggestion carefully costs more than dismissing it and typing. The skill is in fast rejection, not careful evaluation.
Where it will let you down¶
It is fluent about APIs that do not exist. Completion is pattern-matching over plausible code. A method that should exist on a library will be suggested with total confidence. Verify anything you have not used before.
It reproduces your existing mistakes. It learns your file's conventions, including the bad ones. A codebase with a bad pattern gets more of it.
It is weakest exactly where you need it most. Novel logic, unusual constraints, an unfamiliar domain — the cases where you would most value help are the cases with the least pattern to draw on.
Agent mode's confidence is uncalibrated. It will report success on a change that does not compile. Read the diff.
Extensibility is MCP now¶
Since the Extensions shutdown, connecting Copilot to your own tools and data means writing an MCP server. That is a net improvement: the same server works in Claude Code, Cursor and anything else that speaks the protocol, instead of being locked to one host.
MCP is GA in VS Code, JetBrains, Eclipse and Xcode. See Model Context Protocol for how to build one.
The honest picture on productivity¶
Copilot-style tooling is usually sold with a large speedup number. The best available evidence is more equivocal, and worth knowing before you commit a team to a metric.
METR — an independent evaluator, not a vendor — ran a randomised controlled trial with 16 experienced open-source developers on 246 real issues in their own repositories. Developers using AI tools took 19% longer. They had predicted a 24% speedup beforehand, and afterwards still believed they had been sped up by 20%.
A larger 2026 follow-up (57 developers, 800+ tasks, agentic tools rather than autocomplete) moved the point estimate toward neutral — −4% for new participants — with confidence intervals crossing zero, and METR themselves call it "only very weak evidence" because of severe selection effects.
The defensible reading in 2026: there is no credible published evidence that these tools make experienced developers measurably faster on real work in their own codebases, and the perception gap is roughly 40 points wide in the flattering direction. That does not mean they are useless — it means your sense of speedup is not evidence, and you should measure if the answer matters.
Go deeper¶
- Copilot documentation — every surface, and which plan each one needs.
- Cloud agent — the asynchronous agent: a 59-minute session cap and one pull request per task, which shapes what you can hand it.
- METR: measuring the impact of early-2025 AI on experienced developers — the randomised trial where developers were slower with AI and believed they were faster. Read it before quoting any productivity number.
- METR uplift update — the larger follow-up, and a rare case of a group publishing results against its own earlier framing.
Next¶
Claude Code — the other agentic CLI, and a different set of trade-offs.