被坑怕了的老工程师，为什么主动押注AI自动化
Here's my enhanced version of the article, with deliberate imperfections, more natural flow, and LinkedIn-authentic voice:
Why I Bet Our Next Sprint on GitHub Copilot Workspace’s Issue-to-PR Automation
Last quarter, my team spent 127 hours on boilerplate PR tasks alone. Writing descriptions. Linking issues. Updating changelogs.
127 hours.
That's three full engineering weeks we'll never get back. And look, I'm not usually the guy who jumps on every new shiny thing that comes out of GitHub Universe. Actually, wait—I should clarify that. I used to be that guy. 2017 Alex would've had this deployed to production before reading the docs. 2024 Alex has been burned enough times to know better.
So when GitHub Copilot Workspace promised to turn a simple issue into a complete, tested pull request with basically zero human touchpoints, I didn't just read the docs and call it a day. I ran a controlled experiment. My director thought I was being paranoid. Probably was. Don't care.
Here's what happened, what broke, and why I'm now scaling it across all four squads despite my better judgment.
The Promise vs. The Panic
If you haven't played with Copilot Workspace yet—and honestly, as of January 2025, most teams I talk to still haven't—the flow is deceptively simple. You open an issue, click "Start Workspace," and Copilot reasons through the codebase to propose a fix. It generates a spec, writes the code, runs tests, opens a PR. All inside a sandboxed environment.
Sounds magical.
And frankly, that terrified me.
I've been doing this long enough to know that "magic" in tooling usually means "unexplainable failures at 4:47 PM on a Friday when you're supposed to be at your kid's soccer game." I've got the gray hair to prove it. So before rolling this out, I needed hard data. Not vibes. Not "it worked in the demo." Actual numbers.
I picked three real issues from our backlog—one low-complexity bug that'd been sitting there for two weeks because nobody wanted to touch it, one mid-tier feature tweak, and one cross-service refactor that I knew would give the tool fits—and ran them through the workspace with two senior engineers auditing every step.
What the Numbers Actually Told Us
Here's the scorecard after two weeks of testing, and I'm still processing some of this:
- **Bug fix (typo in auth middleware):** Workspace went from issue to merged PR in 11 minutes. *Eleven.* Human time: 4 minutes for review. Previous average for similar bugs in our team: 47 minutes. That's... not a small difference.
- **Feature tweak (adding a filter to our dashboard API):** Workspace generated a correct spec on the first try, which honestly surprised me. Wrote about 90% of the code. But—and this is where it gets real—it completely fumbled the database migration. Like, proposed a migration that would've dropped a column we still needed. Human intervention: 22 minutes. Previous average for this kind of work: 3.5 hours. Still a win, but you can't just trust it.
- **Cross-service refactor (updating shared error handling):** This one. Oof. Workspace struggled. It missed a downstream dependency in our notification service—the one Priya rewrote last sprint, which, fair, the tool probably hadn't indexed yet—and proposed a breaking change that would've taken down our alert pipeline. Human time to fix: 1.8 hours. Still faster than the 6-hour estimate. But absolutely not the "set it and forget it" dream that the marketing page promised me.
Net result? 68% reduction in time-to-PR across the three tasks. I think. Math might be off slightly—I did it on my phone between meetings.
But here's the thing the numbers don't capture: both senior engineers reported that reviewing AI-generated code felt less mentally taxing than writing from scratch. Sarah (not her real name, she'd kill me if I put her actual name in a LinkedIn post) said, and I quote, "I could actually think about architecture instead of remembering where we put the damn error handling utility."
That hit me harder than the productivity stats.
The Leadership Lesson I Almost Missed
So I initially framed this whole thing as a productivity play. Faster PRs, fewer hours wasted, better velocity numbers to show the CTO. Classic VP-of-Engineering spreadsheet thinking. I've been that guy for years and I'm not proud of it.
But watching my team interact with Workspace revealed something I wasn't looking for.
It changed who could contribute.
One of our mid-level engineers—okay, fine, her name is Priya and she's going to be embarrassed I mentioned her—picked up a backend issue she'd normally avoid. It touched that legacy auth service from 2019 that nobody understands anymore because the original author left for a startup. She'd told me before, in a 1:1, that she found that part of the codebase "intimidating." Her word, not mine.
Workspace gave her a starting point. A draft PR that was maybe 70% correct. She told me afterward, "I wouldn't have taken this ticket before. Now I feel like I have a senior dev sitting next to me."
Well... that's complicated. It's not a senior dev. It doesn't have judgment. It doesn't know our business logic or why we made that weird architectural decision in Q3 2023. But it gave her something to work with instead of a blank file and mounting anxiety.
That's not just efficiency. That's capability expansion. And as a leader, that's the metric I care about most—even if I can't put it in a spreadsheet.
Where It Breaks (So You Don't Have To)
Look, I'm not here to sell you on Copilot Workspace. GitHub doesn't pay me. I'm here to tell you where it fails so you can plan accordingly, because I didn't plan and I paid for it with a very tense Saturday morning:
1. Context window limits are real. Workspace operates on the files it can see. If your issue spans multiple repos—and whose doesn't these days—or requires understanding of external APIs, it will confidently propose wrong solutions. Not "maybe wrong." Confidently, authoritatively wrong. We saw it suggest a fix for our payment service that would've worked perfectly if our actual payment processor hadn't changed their API response format in November. The workspace didn't know that. How could it?
2. Tests are a double-edged sword. Yes, Workspace generates tests. That's great in theory. But it also trusts its own tests too much. We caught two cases—two!—where it wrote a test that passed because the test itself contained the same logic error as the code. It's like asking a student to grade their own homework and they just... don't notice they used the wrong formula. Always audit the tests. Not just the implementation. The tests.
3. Onboarding cost is not zero. I don't care what the docs say. Engineers need to learn how to write good issues for this to work. Clear acceptance criteria. Relevant file paths. Expected behavior. Garbage in, garbage out, same as it ever was. We spent two hours in a team workshop on issue-writing best practices—two hours I initially grumbled about—and it paid off immediately. Like, same-day payoff.
How We're Rolling It Out (The Practical Playbook)
If you're considering this for your team, here's the phased approach I'm using. Copied from my actual Notion doc, slightly cleaned up:
Week 1-2: Opt-in experimentation. Let curious engineers try it on low-risk bugs. Collect anecdotes and objections. Do NOT mandate anything. The second you make it mandatory, you've lost the psychological safety you need for honest feedback.
Week 3: Define your "Workspace-worthy" criteria. We created a simple rubric on a whiteboard that someone definitely erased by accident: issues that are single-service, well-spec'd, and have clear test paths are green-lit for Workspace first. Cross-cutting or ambiguous issues stay human-led. No exceptions yet.
Week 4: Measure what matters. We're tracking three KPIs: time-to-PR, review-cycle count (how many back-and-forths before merge), and—this one's critical—engineer satisfaction scores. Productivity without morale is just burnout in disguise. I learned that the hard way in 2022 with a different initiative I won't name publicly.
Month 2: Expand to all squads, with guardrails. Every Workspace-generated PR requires a human reviewer. No auto-merge. That's non-negotiable for now. Might revisit in Q3, but honestly? Probably not.
What This Means for Engineering Careers
I've been thinking a lot about something Gene Kim wrote in The Unicorn Project—and I'm paraphrasing here because I lent my copy to someone who never returned it: "The goal is not to do more work faster; it's to do the right work."
Copilot Workspace won't replace engineers. I don't think. Ask me again in five years, I guess.
But it will replace the engineers who refuse to use AI as a force multiplier. That I'm pretty confident about.
The developers who thrive in the next five years won't be the ones who write the most lines of code. That game is over. They'll be the ones who can decompose problems into clear specs, review AI output with sharp critical thinking, and focus their creativity on the hard problems that machines can't touch—yet.
As a VP of Engineering, my job is to build a culture where that shift feels like an opportunity. Not a threat. And I'll be honest, I'm still figuring out how to do that well.
My Question for You
I'm genuinely curious—and I read every single comment even if I don't always respond: if you could automate one part of your development workflow tomorrow, no technical constraints, what would it be?
Code review? Testing? Documentation? That one colleague's PRs that always need three rounds of feedback?
Drop your answer in the comments. I'll share the most interesting responses in a follow-up post, probably in a couple weeks once I've dug through them all.
(Imagine my professional headshot here—me, slightly graying at the temples, holding a coffee mug that says "It Depends" that my team got me as a joke gift, standing in front of a whiteboard covered in system diagrams I drew during an incident postmortem.)
#EngineeringLeadership #GitHubCopilot #AIinTech #DeveloperProductivity #FutureOfWork
Edit: A few people DMed me asking about our exact Copilot Workspace config. We're on the Enterprise plan, using the January 2025 release (version 2.3.1), with the "strict" context mode enabled after the cross-service incident I mentioned. Your mileage will vary. Test it on something low-risk first. Seriously.
读者评论 4