Detailed Analysis
This Reddit post surfaces a practical workflow tension experienced by a developer using two distinct Anthropic tools—Claude Code and the newer "Cowork"/design-oriented interface—for building products. The user describes a straightforward but consequential problem: when working in Claude's design/cowork mode, context windows fill up rapidly, forcing shorter sessions and more frequent restarts. By contrast, when using Claude Code, the ability to delegate work to subagents keeps the main conversation's context usage low, enabling significantly longer, more sustained sessions. The user is essentially asking whether they're missing a technique to replicate Claude Code's context-management efficiency within the design/cowork environment, or whether this is an inherent architectural limitation of that product surface.
The core issue here touches on one of the most consequential technical constraints in working with large language models: finite context windows. As conversations grow—accumulating file contents, design decisions, iterative feedback, and generated assets—models must hold increasing amounts of information in memory to maintain coherence. Claude Code's subagent architecture addresses this by offloading discrete tasks (e.g., code search, testing, or isolated implementation work) to separate agent instances that process information independently and return only distilled results to the primary session. This keeps the "main thread" lean, since the orchestrating agent doesn't need to retain the full verbose history of every subtask, only the summarized outcomes. Design-oriented interfaces, which are often more visual, iterative, and less modular by nature (involving continuous back-and-forth on aesthetics, layout, and user feedback), may not yet have an equivalent delegation mechanism, meaning all context accumulates linearly in a single session.
This matters because context window management has become a defining differentiator in how effectively AI coding and design assistants can be used for real, sustained product-building work rather than short, disconection-prone tasks. Developers building complex products need tools that can maintain long-running state without constant re-explanation or loss of prior decisions—something that directly affects productivity, cost (since re-establishing context burns tokens), and the practical ceiling on project complexity. The fact that Claude Code's subagent pattern seems to solve this more elegantly than the design/cowork surface suggests an uneven maturity across Anthropic's product lineup: engineering-focused tools have evidently received more sophisticated context-orchestration features than creative/design-focused ones, possibly because coding tasks are more naturally decomposable into modular subtasks than open-ended design work.
More broadly, this reflects an industry-wide trend where AI companies are racing to solve the "long-horizon task" problem—enabling models to work reliably across extended sessions and complex, multi-step projects without degrading or losing coherence. Techniques like subagents, hierarchical memory, retrieval-augmented context, and summarization checkpoints are all emerging as partial solutions. Anthropic, along with competitors like OpenAI and Google, has been investing heavily in agentic architectures that decompose large tasks into smaller delegated units precisely to work around the fixed-size context window bottleneck. This user's experience is a small but telling data point suggesting that as Anthropic expands Claude beyond pure coding into broader "cowork" and design use cases, it will likely need to port over or adapt its subagent/delegation infrastructure to prevent design-focused workflows from hitting the same context ceiling that plagues single-threaded, monolithic chat sessions.
Read original article →