AI writes 60% of Airbnb code, but handoffs are the bigger change

Airbnb CTO Ahmad Al-Dahle says AI now authors 60% of company code, the company shipped nearly 80% more features and improvements year over year and average pull request throughput per engineer rose about 1.6 times. These figures come from his interview with Latent Space and show the scale of an internal shift, not independent proof of software quality.

Code replaced some documents and prototypes shortened handoffs

According to Al-Dahle, Airbnb changed how product, design and engineering teams sequence their work. Instead of a long chain of requirements, Figma design, implementation and testing, teams move to prototypes earlier. Code becomes the central artifact around which they reason together.

The change reaches operations. Customer support was Airbnb's first user-facing AI deployment, and the CTO says agents now resolve roughly half of tickets without a person. The company's second-quarter results reported nearly 45%. Airbnb deliberately keeps safety cases out of automated resolution.

Everest carries one team's experience into the next project

The internal Everest context graph uses LLMs, embeddings and AI retrieval to connect knowledge about the organization and codebase. Airbnb says grocery delivery took 8 to 9 months to build, while the similar airport pickup integration took about 6 weeks. The second team could reuse lessons captured in Everest.

That is a more useful product lesson than the share of AI-authored code. Value appears when a company captures context, runs evals for each use case and selects models by cost, performance and latency. Airbnb operates at least 10 customized models and uses a different trade-off for search than for coding or support.

Generated-code percentages do not count production defects

The 60%, 80% and 1.6 times figures were supplied by the CTO leading the transition. The interview does not publish measurement methodology or a common control group. More pull requests can mean more output, smaller changes or more corrections. Quality remains unresolved without incident, rollback and review-time data.

Al-Dahle names another risk himself: junior engineers may lose some of the experience from which judgment develops. Airbnb therefore requires every engineer to explain the code in a pull request even when AI generated it.

On-call agents will test whether context becomes accountability

Airbnb is beginning to use asynchronous agents in containers, triggered by monitoring events. An agent can triage an incident, propose a pull request or close a flaky alert. This is where Everest and use-case evals face higher-stakes work.

The next metrics should include incidents caused by AI changes, human review time, rejected pull requests and support outcomes by problem type. Throughput growth only matters when the repair queue does not grow behind it.

Lilith's verdict

Airbnb has placed AI directly between prototype and production, and 60% is only the number on the door. Inside, the test is whether an engineer can explain every pull request and the overnight on-call agent leaves a readable trail.