AI Agents Won't Fix Your Shipping Bottleneck

Dax Raad, co-founder of the explosively popular open-source coding harness OpenCode, delivers a startlingly honest assessment that cuts against the AI hype cycle. Despite building a tool used by millions of developers and operating in the aggressively competitive coding agent space, Raad admits a core truth: AI hasn't made his team ship better software faster. Objectively, individual coding tasks have become easier, but the hard parts—determining what to build, maintaining product cohesion, and strategic thinking—are untouched by current AI.

The real bottleneck shifted from writing code to managing the explosion of poorly considered features that AI makes trivial to generate. Raad describes a “Frankenstein” product risk where shipping ten times as many features doesn't equate to ten times as many good ideas, creating lasting maintenance burdens. This led him to write a memo to his team explicitly trying to slow them down and discourage mindless one-to-one AI prompting for every user request or competitor feature. His experience confirms a pattern: AI inflates quantity of output but doesn't solve the synthesis and taste problem that defines great software.

Raad highlights a critical absence of competitive advantage, even among coding agent companies themselves. No player in his space is “crushing” competitors through superior AI use, despite universal AI adoption. If AI conferred a decisive ship-velocity edge, it would be most visible in the coding agent market itself—but it isn't. This suggests current tools primarily improve individual developer happiness and workflow feel rather than organizational throughput, echoing a provocative idea that the same work gets done, but engineers are happier.

Operationally, Raad’s strategic thinking emphasizes positioning over raw technical superiority. OpenCode’s success stems from reading industry dynamics: claiming the open-source, model-neutral territory, then letting competition between frontier model providers—Anthropic, OpenAI, and others—drive adoption through conflict. The old playbook of galvanizing competitors against one “temporary bad guy” continues to work, allowing a small team to leverage billion-dollar companies’ own rivalry. This positioning insight is more durable than any specific AI feature.

On the business side, inference is revealed as a surprisingly high-margin operation, with some models showing up to 80% margins even for a company renting middleman GPUs. Raad notes GPU supply is tightening dangerously as demand scales faster than hardware production, creating a bottleneck that affects even fast-growing startups. Enterprise adoption demands unglamorous but essential control planes—SSO, budget controls, permissions—that are the real monetization path for tools like OpenCode.