Mikyo King describes code mode as an alternative to supplying an agent with a large catalogue of tools and inserting every intermediate result into its context. An agent can find relevant tools and compose calls in codeAgent tool use lets an agent call bounded external capabilities so it can retrieve information or change state beyond generating a response., returning only the information neededContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task. for its task.
In Arize Phoenix, King says the team uses Bash, a Python sandboxAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access., SQL and a browser-bound programming interface to let its agent combine actions and query observability data. He argues that this approach allows agents to perform longer tasks and create useful operations dynamically.
King also identifies trade-offs: sandbox resources, network egress, secrets and permissions need careful controls, while failed code can be harder to debug. Durability, nested sandboxes and the capability of smaller models remain open questions.
For Arize’s own Harbor-based evaluationEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria., King reports a 43% reduction in token use and a 31% reduction in cost after adopting code mode, with latency remaining broadly constant. These results describe the evaluated Arize workloads, not a general guarantee.
Watch on YouTube
