Modern AI workflows are shifting from conversational chat toward integrated 'harnesses'—specialized interfaces and tools designed to execute tasks rather than just answer questions. Users are increasingly adopting tools like Claude Code, Codex, and OpenCode to manage the heavy lifting of software development. These harnesses often provide more ergonomic experiences than standard web interfaces, allowing for better integration with existing development environments like VS Code or Emacs. For many, the goal is to move from manual prompting to a 'set and forget' model where the tool handles the execution of sub-tasks, such as creating documentation or running tests, while the user focuses on high-level architecture.
A recurring pattern in these advanced setups is the use of a 'coordinator' agent to manage a swarm of sub-agents. In this architecture, a primary model (often Claude) acts as a central hub that receives a high-level request and delegates specific actions to a group of sub-agents (such as Codex). This approach simplifies the user experience by providing a single point of contact for complex tasks, while the coordinator handles the internal communication, verification, and sequencing of the sub-agents. This method is particularly effective for large-scale data analysis or complex coding projects where a single model might struggle to maintain context across many disparate tasks simultaneously.
Despite the push toward sophisticated orchestration, there is a clear trade-off between out-of-the-box convenience and custom-built infrastructure. While many users find that standard tools like Claude Code provide the most cutting-edge agentic features, others are building bespoke tools—such as custom kanban-style management systems or specialized research swarms—to handle niche requirements like equity analysis or personal assistant tasks. These custom setups often involve a mix of high-tier subscriptions, pay-as-you-go APIs from providers like OpenRouter, and even local inference on specialized hardware. The choice between a standard harness and a custom swarm often depends on whether the user's needs are sufficiently unique to justify the overhead of building and maintaining their own orchestration logic.