Skip to content
#Agents

[1/2] Agentic coding at Airbnb (Agents column)

#Agents #Engineering #AI #Metrics #Software #DevEx #Productivity #DevOps #Architecture #Culture #ML #SystemDesign

It's another one. note on September DPE SummitThe one I talked about earlier. In it, the guys from Airnbnb told how they built an agency model of development and moved from classic PDLC to AI-native development. (approximately in the direction I described in eponymous). And here we are not talking about some toy applications created by agents, but about agency development in their monorepe with standard requirements for security, compliance and with 1k+ engineers in it. Moreover, the authors used a metaphor from pair programming, XP practice to describe the approach. (extreme programming)There were two engineers doing features together: one was a navigator thinking about strategy, and the other was a pilot holding a keyboard and writing code. In this metaphor, the role of the pilot was taken by the agent, and the role of the navigator was left to the engineer. In the beginning 2025 In the past few years, I have been expecting this regime to be extended to 20% - 40By September, they had already exceeded the plan and reached approximately 60%

Airbnb’s path to this is predictable and understandable.

  1. At first, they had an IDE-first approach: internal plugins with good UX, access to editor context, and a proprietary knowledge/RAG Pipeline for Airbnb specifics.
  2. Then the market moved sharply towards CLI-agents: they were stronger as agency tools, were able to take several steps in a row, call tools and make local decisions.
  3. Airbnb adapted and tried to look at the world as a CLI-first, but quickly saw the problem: external CLI tools gave cool answers for external projects, but did not know their monorepees, internal patterns and code base of the company.
  4. Then they made a very mature move: they did not argue with the IDE or CLI, and assembled a small core team and cross-functional working group to bring all surfaces to the same level. The team had just four engineers plus product support, so they deliberately narrowed the focus to three bets: Bring agents to Airbnb Bring Airbnb knowledge to agents Standardize around MCP In parallel, they deployed the champion/community model through hackathons, talks and an internal working group.

In continuation I'm going to talk about how the guys went about achieving those goals.

#Engineering #AI #Metrics #Software #DevEx #Productivity #DevOps #Architecture #Culture #Engineering #ML #SystemDesign