Pinned Tweet
Greptile has raised a $50M Series B led by Menlo Ventures. The round values the company at $750M.
Greptile reviews and tests more than 1% of all PRs opened on GitHub every day. Thousands of engineering teams including Nvidia, Coinbase, Decagon, and Vanta use Greptile as their validation layer.
Fewer than 50 people, we are a small, talent-dense team of engineers and researchers in a quiet corner of San Francisco. If you’re interested in joining us, we’d love to hear from you.
Greptile reviews and tests more than 1% of all PRs opened on GitHub every day. Thousands of engineering teams including Nvidia, Coinbase, Decagon, and Vanta use Greptile as their validation layer.
Fewer than 50 people, we are a small, talent-dense team of engineers and researchers in a quiet corner of San Francisco. If you’re interested in joining us, we’d love to hear from you.
128
54
895
261,878
At Greptile, we are able to ship fast because of the way we experiment with our agents. Models, harnesses and best practices keep changing, and over the past year, the pace has only accelerated. We've always had a backlog of new tools to try, and prompts to tweak, so we needed a quick way to test them.
So we built what we call the experiments meta-harness. Every agent setup lives in a template file that declares which models and harness it uses, what environments it runs in, which tools and skills it can access, and which prompt version it follows. These files sit in a central registry and are loaded during a Greptile review.
Trying a new model or prompt is as simple as adding a new file. We've since added custom workflows, gradual rollouts, automatic analysis of agent traces, and recommendations for what to test next. We've been able to run thousands of experiments this way to further understand which models and tools help (or hurt) a code review.
Every agent-powered flow at Greptile builds on top of this now. We pick a goal, our researchers and engineers decide how to measure it, and we see if we can beat the current best setup. Measuring here is the hard part. Most code review benchmarks test whether a reviewer can find bugs in a curated set of examples, and this tells you little about what customers actually care about. Customers want to know if the comments are correct, if they are worth reading and if the issue gets fixed. So we score every experiment on our own evals, built around these three questions.
The speed at which we can test new combinations of tools decides how fast the product gets better. And with a system able to analyze its own traces and suggest what to try next, it's our first step toward our own version of recursive self-improvement.
So we built what we call the experiments meta-harness. Every agent setup lives in a template file that declares which models and harness it uses, what environments it runs in, which tools and skills it can access, and which prompt version it follows. These files sit in a central registry and are loaded during a Greptile review.
Trying a new model or prompt is as simple as adding a new file. We've since added custom workflows, gradual rollouts, automatic analysis of agent traces, and recommendations for what to test next. We've been able to run thousands of experiments this way to further understand which models and tools help (or hurt) a code review.
Every agent-powered flow at Greptile builds on top of this now. We pick a goal, our researchers and engineers decide how to measure it, and we see if we can beat the current best setup. Measuring here is the hard part. Most code review benchmarks test whether a reviewer can find bugs in a curated set of examples, and this tells you little about what customers actually care about. Customers want to know if the comments are correct, if they are worth reading and if the issue gets fixed. So we score every experiment on our own evals, built around these three questions.
The speed at which we can test new combinations of tools decides how fast the product gets better. And with a system able to analyze its own traces and suggest what to try next, it's our first step toward our own version of recursive self-improvement.
5
3
49
2,757
Validating every piece of software on the planet is going to take a lot of tokens.
Thanks @togethercompute for making it fast 💚
Thanks @togethercompute for making it fast 💚
greptile wants to validate every piece of software on the planet
as more code gets written by ai, the first review pass needs to be independent + fast. greptile’s v5 agent spins up sub-agents on every pr, pairing open + frontier models so one model’s blind spots don’t decide her ai powers part of the inference behind those open models, helping reviews come back faster and at lower cost
watch how @greptile does it
1
1
15
1,653
thank you @Nasdaq, very cool
Greptile has raised a $50M Series B led by Menlo Ventures. The round values the company at $750M.
Greptile reviews and tests more than 1% of all PRs opened on GitHub every day. Thousands of engineering teams including Nvidia, Coinbase, Decagon, and Vanta use Greptile as their validation layer.
Fewer than 50 people, we are a small, talent-dense team of engineers and researchers in a quiet corner of San Francisco. If you’re interested in joining us, we’d love to hear from you.
2
5
65
2,416
We're getting the blackboard out again on November 4th. Here's who's teaching this time:
@aidenybai on how React actually works, rebuilt from scratch
@ficus on what microwaves, breakers, and Android permissions have in common
@rohankumar on why quantum computing has been “ten
Doors open at 6:30pm. Drinking in class is encouraged.
@aidenybai on how React actually works, rebuilt from scratch
@ficus on what microwaves, breakers, and Android permissions have in common
@rohankumar on why quantum computing has been “ten
Doors open at 6:30pm. Drinking in class is encouraged.
2
3
28
3,729
Greptile retweeted
i gave a talk at @aiDotEngineer yesterday called "how to build a code review agent"
it is now published in full on our blog.
some teams want to build their own ai code review agent instead of using greptile.
we help them to do that because we know a thing or two about how to n a lot in the process.
this post covers our most useful advice on how to build a great code review agent:
it is now published in full on our blog.
some teams want to build their own ai code review agent instead of using greptile.
we help them to do that because we know a thing or two about how to n a lot in the process.
this post covers our most useful advice on how to build a great code review agent:
15
7
113
7,589
We love open source and believe the companies that benefit from it should help sustain it. We're honored to back @omarchy and the Omacom Foundation with $150,000 in tokens for code review.
Please welcome @greptile as a Distinguished Corporate Patron to the Omacom Foundation, as they've dedicated $150,000 in tokens to continue providing PR reviews across all our repos. Agentic peer review that has already delivered! omarchy.org/news/2026/10/g…
3
3
59
4,487
Greptile retweeted
running your own internal greptile is usually more expensive than paying for greptile.
usually $5-$20/review to get in the vicinity of greptile quality vs $1/review for greptile.
we have one of the largest dataset of buggy PRs in the world. this means we can know for sure if a particular version of our agent is more or less effective than another.
as a result, we can experiment with hundreds of permutations of models, harnesses, techniques, tools to eek out maximum performance per dollar.
we then pass those savings onto the user.
we can only do this because of the dataset. if we didn’t have it, we won’t know for sure if a new version really can catch every unique one in a million bug.
usually $5-$20/review to get in the vicinity of greptile quality vs $1/review for greptile.
we have one of the largest dataset of buggy PRs in the world. this means we can know for sure if a particular version of our agent is more or less effective than another.
as a result, we can experiment with hundreds of permutations of models, harnesses, techniques, tools to eek out maximum performance per dollar.
we then pass those savings onto the user.
we can only do this because of the dataset. if we didn’t have it, we won’t know for sure if a new version really can catch every unique one in a million bug.
after doing a bunch of work to improve Aether's PR review feature I can definitely say I underestimated how difficult it is to build a system of reviewing code that's fast, accurate, and relatively cheap
good job @dakshgup
28
9
257
67,399
We're at AI Engineer NYC this week!
Here's how you can find us:
→ Booth #27, Monday to Wednesday
→ Tue 2:55pm ET: @dakshgup on how to build an AI code review agent
→ Tue night: rooftop happy hour + fireside chat with Daksh and @nerdsane at @datadoghq HQ
See you there!
Here's how you can find us:
→ Booth #27, Monday to Wednesday
→ Tue 2:55pm ET: @dakshgup on how to build an AI code review agent
→ Tue night: rooftop happy hour + fireside chat with Daksh and @nerdsane at @datadoghq HQ
See you there!
2
3
16
1,909