Part 3 of 3 — Agentic AI Success Stories

The pull request is the product.

Part 2 put a human at the start, before anything was written. Coding agents put a human at the end: merge. Where agents actually make sense already named automated tests (CI) and code review as the checks that decide success. What’s coming next named autonomous software development as a pattern. This part is one real product, not a Copilot roundup on that page.

Roundups still say “Copilot Workspace.” The live GitHub product to cite is the coding agent (GitHub also documents it as Copilot cloud agent): you assign work, it opens a pull request, and it asks for review.

GitHub Copilot coding agent

GitHub’s 19 May 2025 product post: assign a GitHub issue to Copilot (or prompt it). It spins up an environment on GitHub Actions, pushes to a draft pull request, and you track it in session logs. Human approval is required before CI/CD workflows run on that work. Branch protections still apply. The person who asked for the PR is not the one who can satisfy a required-review rule by approving it themselves.

The 25 September 2025 changelog marked the coding agent generally available for paid Copilot subscribers: delegate a task, draft PR in the background, request a review, iterate via PR comments.

GitHub’s docs say the same thing: assign the issue, Copilot raises a pull request, then requests your review. Workflows do not run automatically on the agent’s pushes until someone inspects the diff — especially anything under .github/workflows/.

issue or prompt → agent branch → draft PR → human review → merge
                                      │
                              CI waits on approval

That is the gate. Judgement calls, such as “should we ship this?”, stay with a person. Prompts still matter: a vague issue is a vague PR. You still need a spec, a stop, and a path to a person.

Stripe — a vendor story, their number

Anthropic’s Stripe customer page says Claude Code was deployed to 1,370 engineers, pre-installed, with a signed enterprise binary. One cited result: 10,000 lines of Scala migrated to Java in four days, estimated at ten engineering weeks by hand. Scott MacVicar’s framing on that page: treat the assistant as a capable new engineer that needs context, not a replacement that works on its own. That is a vendor story, dated to that page — not a comparison against Copilot.

Why AI Agents Fail covers what happens when you skip the review. This part only says the success story is the review.

Takeaway: A coding agent that cannot open a gated PR is a demo. One that waits for merge is the real job.

Sources