Amol Kapoor

Back to table of contents.

A Founder's Guide: Collaboration at Google

Written Feb, 2022. Published March, 2022.

Collaboration is "the action of working with someone to produce or create something". This is a surprisingly nuanced definition.

First, collaboration is an "action". It's not something that happens passively, it is something that requires an actual change from the default, which in turn implies thought and planning.

Second, collaboration involves "working with someone". The word 'with' has a lot of definitions, it's something you probably don't ever even think about because it's so common but Google says there are 10 unique definitions for the word. The most relevant is 'in the same direction as'. You are working 'in the same direction as' someone else. This implies alignment — of goals, values, techniques. You cannot collaborate without being aligned.

And finally, collaboration requires the 'production or creation' of something. You don't collaborate in a vacuum. Collaboration has an end state, a goal in mind.

Collaboration, at some level, is baked into the DNA of every company. You start a company — and hire other people and so on — because it's more efficient than working on your own. But collaboration is really fucking hard. Raise your hand if you've heard this story: a bunch of people get together to do a thing, everyone works on their own version of the idea until there's overlap and clash, nothing is agreed upon and the group falls apart without accomplishing anything. You know what's not hard? For people who are nominally collaborating to actually be working in isolated silos.

The joke writes itself.

Why does this happen? Well, first, because collaboration depends on communication, and we already discussed how painful that is. But also, people can't read minds. Even the best intentioned developers cannot fully understand what another developer is doing without collaborating in some way. Software is just too complex, too abstract. So if there isn't a serious attempt to regularly sync on some mutual understanding, people diverge until you one day realize none of the pieces are actually compatible.

But that's in the best case. In practice, many developers don't like collaborating. We all have our own styles of programming, our own way of doing things. Programming is an expression of thought, and trying to get into the thought process of another person is frustrating and time consuming and, frankly, sometimes irrationally deeply upsetting (see: tabs and spacing, single vs double quotes, 80ch line limits…). So developers naturally isolate themselves.

From a managerial and organizational perspective, this is bad. Isolation means bottlenecks. It means dependencies on individual people or teams. And most importantly, it slows down development corkscrews — if you aren't collaborating, you have no reason to make your code readable or usable for more general use cases. But you can't force people to collaborate, not really, because collaboration happens at every level of the company, from hundred person orgs all the way down to individuals, during public all hands and 'in the gaps' when no one is looking. So you have to instead make collaboration something that is worthwhile, such that every person is naturally incentivized to collaborate by default.

At Google...

At Google, collaboration was mostly the default. I've already written about Google's efforts to make communication seamless. I think Google took that foundation and did a great job setting up systems that would encourage people to seek out others and work together. For the most part, these systems functioned by implicitly or explicitly tying collaboration to personal achievement metrics.

Now, before going further, I have to note that many even of these systems felt frustrating for many developers. The organizational tools below are often maligned for being stressful and difficult and unenjoyable and inefficient. I'm not surprised. Collaboration is hard, and often uncomfortable, so managerial levers that encourage collaboration are looked upon with distrust and skepticism. Nonetheless, I think the net effect was overwhelmingly positive for Google's forward mobility.

With that out of the way, let's talk about why perf is great.

Perf and Peer Review

For those of you who have luckily avoided being within earshot of a Googler in February, March, August, or September, Perf is shorthand for Performance Review. It is a bi-annual (meaning twice a year) review period of about a month, where engineers are asked to write somewhere between 1000 and 5000 words on what they did and how awesome they are. Perf is used as the primary input for calculating bonuses and, eventually, promotions. And since every promotion relies on all past Perf results, there is a strong incentive to write a 'good' Perf every cycle.

This is very stressful.

But I'm not here to talk about stress, I'm here to talk about a key aspect of Perf: the peer review. See, in order to have a 'good' Perf, you need to solicit short and long form reviews from peers. This feedback is heavily counted when evaluating an employee's trajectory. Importantly, reviews from high-level folks outside your team are weighted higher than internal reviews — if your director thinks you're great, that's whatever, but if a different director thinks you're great, that's a big fucking deal. This is especially true as you go up the promotion ladder.

Naturally, this system inventivizes anyone who wants a good Perf (i.e. everyone) to constantly think about how they can interact with and provide value to other people on other teams. That in turn encourages good API design, clear documentation, and generalized multi-purpose tooling (see corkscrew development). Critically, this works both ways. If I help you, we both can give each other strong peer reviews. So we end up having a very good reason to align on goals and strategy, and an equally good reason to develop a process that is legible -- enough that we could write about it later.

Peer review happens at every level of the command hierarchy at Google. As a result, there is an organic bottom-up drive to work across teams and orgs.

Maybe not coincidentally, teams that promoted candidates without strong peer reviews were, in my experience, harder to work with. They were also much more likely to be re-orged or disbanded.

Impact metrics

Any conversation about Perf requires mentioning 'impact'. At Google, 'impact' is a catch-all term that means 'how much did my work change the company'. Most of the time, impact is measured in dollars. But it can also be measured in clicks, or latency, or engagement time. Broadly speaking, the more impact, the more likely you are to get promoted. So people cared a lot about impact.

You know what else counts as impact? Speeding up your coworkers. If you can make some task faster for your peers, you can claim impact — and get credit for some of the downstream impact that they showed. When I was at Google, the joke was that if you wanted a fast promotion, you should go work on internal infra (like Borg). It's easy to rack up impact when every code change touches every engineer at the company. But Google isn't stingy, even good tutorials count as impact.

This accounting system turns every engineer into a devops person. In practice, I think this policy really helps smooth things out on the margins. Every time an eng builds a handy bash tool, or works through some tricky configuration, they might spend the extra day to write and publicize some good docs on what they did and why. The usual corkscrews apply.

Generally, these sorts of libraries or utilities are win-win. Everyone is looking for impact. Since speeding up other people genuinely helps them achieve more, there is strong incentive to seek out and re-use tools other people build. So for any given utility, the developer wants as many people as possible to use it; and anyone who can benefit is already seeking it out.

Peer Bonuses

The last piece worth highlighting in Google's arsenal of incentives is the peer bonus program. A peer bonus is a $200 bonus on your paycheck. Every Google employee is given five peer bonuses that they can hand out each quarter. You can't give them to anyone in your direct management chain, nor anyone who's given you a peer bonus that same quarter. They don't roll over, and you don't benefit from keeping them, so there's no reason not to use them. Generally, the recipient's manager is alerted and everyone on their team reaches out to congratulate.

Peer bonuses are amazing.

Peer bonuses are a great way to encourage folks to go the extra mile. They are a natural, decentralized mechanism that rewards kindness and helpfulness. And receiving a peer bonus really does make you feel like a million bucks, even if the actual change to your paycheck is minimal. I think peer bonuses send a strong cultural signal about what is important from top down, in a way that really impacts engineering behavior in all the subtle ways that are necessary for a good collaborative environment.

(O, and peer bonuses are a great way to rapidly identify who on a team is carrying the weight or providing a lot of value.)

Closing thoughts

This ended up being a lot longer than I thought it would be, but I think that's because it is hard to incentivize collaboration. But it's also an incredibly critical thing to manage, whether you're building a startup or any other kind of organization. It's an active process, and therefore requires active attention and explicit supporting policies.

Google has 100k employees. Yes, there are conflicts, and yes, teams aren't always on the same wavelength. But I worked with folks from every part of Google — from Photos to Robotics to Ads to YouTube to Health — and because every person I met was incentivized to collaborate we were able to accomplish some pretty awesome things. I don't think they were doing anything special, though, and the policies they put in place are easy to copy. Once you nail it down, you can easily 100x your organization's productivity.