Inside Airbnb’s AI Transformation: How CTO Ahmad Al-Dahle Rebuilt Engineering Workflows

Central to this workflow is an internal organizational context graph called Everest, which uses large language models, embeddings, and AI retrieval to help developers navigate the codebase. By referencing context stored during previous projects, Everest allows generalist engineers to work on specialized code segments. For instance, while building a grocery delivery feature required eight to nine months, leveraging Everest allowed Airbnb to build a similar airport pickup service in about six weeks. CEO Brian Chesky highlighted both services during the company's 2026 Summer Release in May.

Multi-Model Infrastructure and Asynchronous Agents

Airbnb operates as a multi-model company, deploying over 10 customized models in production. Rather than relying on a single system, the company evaluates models along a Pareto frontier based on cost, latency, and performance. Coding tasks prioritize accuracy and tolerate higher latency, making high-capability frontier models ideal despite higher costs. Conversely, search functions require low latency at scale, leading Airbnb to use smaller, specialized models fine-tuned through post-training and reinforcement learning.

In user-facing operations, AI agents handle approximately 50% of customer support queries. Before reaching production, these systems are tested against synthetic data batteries generated specifically to evaluate edge cases and safety constraints. Internally, Airbnb utilizes an organizational agent called AirChat, which incorporates Model Context Protocol (MCP) data. The company is also moving toward asynchronous agents running in containers that trigger automatically based on monitoring events—such as Grafana alerts—to triage on-call incidents, propose pull requests, or close flaky alerts.

For developers looking to implement multi-model architectures similar to Airbnb's setup, testing top AI models cheaply through one API is possible via https://apixoai.online.

What it means for developers

Airbnb’s implementation provides several concrete lessons for software engineering teams adapting to AI-driven workflows:

1. Focus on Outcomes Over Features: Al-Dahle noted that organizing teams around missions and measurable results is more effective in an AI-driven environment than structuring teams strictly around feature ownership.
2. Pragmatic Model Selection: Matching specific workloads to appropriate models based on cost and latency profiles yields better performance. Smaller post-trained open models can outperform frontier models on narrow, high-volume tasks like search, while frontier models remain essential for complex, error-sensitive code generation.
3. Synthetic Evaluation: Deploying autonomous agents safely requires robust synthetic data evaluation pipelines prior to live production releases, particularly for high-stakes environments like customer support.
4. Maintaining Core Engineering Craft: A key concern raised by Al-Dahle is ensuring junior developers build foundational engineering judgment when AI handles much of the writing. Airbnb addresses this by enforcing a policy where every engineer must be capable of explaining any AI-generated code or pull request, ensuring developers retain understanding of software architecture, testing, and interface design.

Source: Inside-Out AI: Rebuilding Airbnb Behind the Scenes and Across the Guest Experience — Latent Space. Written by the Apixo team from that report.