How Briya uses CodeRabbit to govern every AI coding agent
Briya uses CodeRabbit as an independent review layer across coding agents and repositories, turning engineering conventions into automated standards while keeping humans in control.
CASE STUDYAbout 60% of CodeRabbit suggestions accepted
1,000+ Linear MCP checks and counting
Engineering conventions encoded once and enforced across repositories
Human approval preserved for architecture, business logic, and compliance
Briya builds a healthcare-specific AI platform that supports researchers from hypothesis to insight. Its technology connects to on-premises data, harmonizes patient information inside secure environments, and lets medical researchers query complex datasets in natural language without writing SQL. With about 50 employees across New York and Tel Aviv, Briya recently opened its browser-based AIRE SaaS platform to the public at no cost.
AI coding agents helped the team move faster. They also made code review the bottleneck. As output increased, engineers experimented with different tools, each with its own blind spots, while standards that had never been written down began to drift.
Writing software has become a much faster thing than reading the code. We've been producing so much code, and code reviews have now become the internal bottleneck.
— Elad Itzhakian, Executive Director of R&D at Briya
Challenge: Standards that survive the tool being used today
Briya did not have a problem with one particular coding agent. It faced a structural problem across AI coding tools: large language models are probabilistic, different tools catch different errors, and a tool can develop tunnel vision around its own output.
An independent review layer gives every pull request a fresh read with repository context beyond the diff. But that independence only works when engineering standards are not locked inside whichever tool wrote the code. If learnings live inside one coding agent, switching agents means starting over.
Briya's previous review tool added another constraint. It worked repository by repository and could not propagate standards across a multi-repo codebase. Conventions that were not explicitly documented stayed in engineers' heads and were enforced inconsistently as the team grew.
LLMs are flawed by design. They are non-deterministic, probabilistic tools. You should expect some error rate. The way the world has chosen to deal with it is having second, third, fourth opinions. Different tools find different errors. You just try again, from different angles, with different contexts.
— Elad Itzhakian
For Briya, this was not a tooling-preference problem. It was a governance problem, and the answer was automation.
I'm not a big believer in spoken conventions. I'm more into automation.
— Elad Itzhakian
~60%
suggestion acceptance rate
1,000+
Linear MCP checks
1
governance layer across coding agents
Why CodeRabbit
Learnings turn conventions into automation
CodeRabbit Learnings let Briya automate rules that used to live in engineers' heads, including avoiding foreign keys in migrations, enforcing logging, and flagging recurring issues. The team can see which rules matter most based on how often each learning fires, turning institutional knowledge into standards that scale.
Multi-repo analysis puts standards everywhere
CodeRabbit links repositories so standards propagate automatically. Engineers do not have to memorize the rules for every repository, and conventions follow the code wherever it lives.
Full-codebase context catches what experienced reviewers miss
CodeRabbit reviews the full codebase, not only the pull request diff. It catches cross-file inconsistencies, parameters dropped from lengthy queries, and changes that diverge from established patterns elsewhere in the repository. That context helped win over skeptical engineers quickly.
It takes maybe one or two comments where an engineer says, 'wow, that's correct, I would not have paid attention to this.' From that point, they're in.
— Elad Itzhakian
High-signal feedback earns trust
In healthcare, a noisy review tool loses trust quickly. Briya's engineers accept about 60% of CodeRabbit's suggestions, meaning the team acts on most of what it flags instead of dismissing the feedback as noise.
As soon as a tool starts making noise, engineers say, 'I can't trust this.' It loses them very quickly.
— Elad Itzhakian
Linear MCP verifies intent, not only code
Briya connected CodeRabbit to Linear MCP so every pull request can be checked against its Linear ticket. The review verifies that the implementation matches the original intent, not merely that the code looks correct in isolation.
Since starting its trial in May 2026, Briya has completed more than 1,000 Linear MCP checks. The workflow mirrors how the company uses coding agents, which are also instructed to check their work against the ticket first.
Results: Briya shipped faster and kept humans in control
CodeRabbit became the review layer above every coding agent Briya used. The team could try new tools, switch agents, or combine approaches without losing its standards. With Learnings, conventions are encoded once and enforced everywhere.
You have to face the fact that every tool in this domain is new and people want to try around and see what works best. Presenting CodeRabbit as a different layer that doesn't necessarily conflict with other layers, that's a good direction.
— Elad Itzhakian
A compliant, human-in-the-loop workflow without the bottleneck
Briya kept its single-reviewer policy for compliance. CodeRabbit handled the first-pass details so human reviewers could focus on architecture, business logic, and regulatory considerations.
That division of work helped the team move quickly while preserving human approval for the decisions that required judgment.
It is safe to say that CodeRabbit has played a part in our ability to ship it fast.
— Elad Itzhakian
CodeRabbit = One governance layer for every coding agent
Before CodeRabbit
- AI coding agents generated code faster than humans could review it.
- Engineering conventions lived in people's heads and drifted across tools.
- Standards could not be enforced consistently across multiple repositories and agents.
- Switching or combining agents risked losing context and accumulated learnings.
With CodeRabbit
- About 60% of suggestions are accepted, giving engineers high-signal feedback they trust.
- More than 1,000 Linear MCP checks have validated pull requests against ticket intent.
- Learnings encode conventions once and enforce them automatically across repositories.
- Human reviewers stay focused on architecture, business logic, and compliance.
The bigger picture: Governing every AI coding agent
No single coding agent catches everything. Briya's approach is to use multiple tools, enforce consistent standards across all of them, and make CodeRabbit the independent layer that holds the workflow together.
It only makes sense for AI to read the code that other AIs generate and have a human approve when needed.
— Elad Itzhakian
Code reviews have become a major bottleneck in fast-paced teams. Having something like CodeRabbit helped us understand that it's not a bottleneck you just have to live with. There is something you can do about it.
— Elad Itzhakian, Executive Director of R&D at Briya
New York, USA and Tel Aviv, Israel
https://briya.com/Industry
Healthcare AI and clinical research
Challenge
AI coding agents accelerated delivery faster than human review capacity, while standards drifted across tools and repositories.
