Dominik Kundel traces a request through two open interfaces: the Codex app server connects a user interface to the harnessAn AI agent harness is the software framework that packages a model with tools, instructions, context management, execution controls, and user interaction., while the Responses API connects the harness to model inference. He presents both as parts of an open ecosystem that lets developers build custom interfaces, harnesses and compatible model-provider integrations.
Dominik Kundel describes context constructionContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task. as a balance among size, flexibility, performance and cacheability. He shows how deferred tools remain discoverable through tool search without occupying the initial context and how the available-skills list is constrained as a proportion of the model's context window.
Dominik Kundel explains how the harness handles asynchronous subagentsAgent orchestration coordinates AI agents, tools, people, tasks, state, and control flow so a larger workflow reaches a verified outcome. and background terminals, drives browser interactions through code in a persistent JavaScript environment, edits files with patches and uses shell tools for search and navigation. He also outlines separate sandbox implementationsAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access. for macOS, Linux and Windows.
Dominik Kundel closes with mechanisms for longer and safer agent work. A read-only review agent evaluates risky actions against user authorization and impact, persistent WebSockets reduce repeated network payloads, a goal loop injects the objective until completion is reported, and server-side compaction carries essential context into subsequent turns.
Watch on YouTube
