Expedia Group's engineering team shares lessons from integrating AI agents into their planning and development workflow, arguing that missing context doesn't slow agents down the way it slows humans down — instead it accelerates confidently wrong work. The team built a context repository combining markdown proposal documents (for humans) with a machine-readable STATUS.yaml file (for agents and scripts) tracking stage, priority, affected services, and OKR alignment. They also adopted Anthropic's 'agent skills' pattern (SKILL.md files) to make context-checking repeatable. A concrete win came from their Model Monitoring initiative, where an agent compressed research across nine vendor alternatives, surfacing licensing risks like Deepchecks' AGPL 3.0 and Seldon Alibi Detect's BSL 1.1. The unsolved problem remains 'context drift' — decisions made informally in Slack or hallway conversations that never make it into the structured context, leaving agents reasoning from stale information.
Table of contents
Expedia Group Technology — EngineeringMissing context used to slow a team down, with agents in the loop it speeds up the wrong work insteadThe problem was hidden contextThis is not just more documentationGet A Talhan’s stories in your inboxHow it changed the workWhat happened when we used itWhat is still hard: context driftStart with one workflowQuestions this post answers
What is an agent skill file (SKILL.md) and how does it work with AI coding agents?
An agent skill is a folder containing a SKILL.md file with YAML frontmatter specifying a name and description, followed by a procedure written in plain markdown. An agent loads only the name and description of every installed skill upfront, and reads the full procedure body only when it judges the skill relevant to the current task, making the description function as the trigger condition. Teams standardizing agent workflows can follow ongoing coverage of agent skill patterns on daily.dev.
Why do AI coding agents produce confidently wrong output when project context is incomplete, unlike human engineers?
A human engineer missing context typically slows down, asking someone or searching old notes, but an agent tends to keep moving and produce work that is technically plausible and well formatted while being aimed at the wrong reality. This happens because unclear context does not just delay agentic work, it accelerates the wrong work, since the agent fills gaps with assumptions rather than pausing. Developers tuning agent-in-the-loop workflows can track lessons like this on daily.dev.
What licensing risks came up when comparing model monitoring vendors like Deepchecks and Seldon Alibi Detect?
Deepchecks carried an AGPL 3.0 licensing risk, while Seldon Alibi Detect used a BSL 1.1 license that requires care before production use. These licensing distinctions emerged during a structured comparison of nine model monitoring alternatives evaluated across drift detection, data quality, evaluation, explainability, scalability, integration, and cost. Anyone evaluating monitoring tools under different open-source licenses can follow related comparisons on daily.dev.