OpenAI’s Codex-based models focus on keeping one continuous thread alive, using server-side compaction to squeeze long conversations into a smaller context window. When the token count creeps up, Codex automatically summarizes or filters out tool calls behind the scenes, then stitches the core information back into the active prompt. That way you get coherence across hundreds of thousands of tokens without juggling multiple agents. The trade-off: everything runs through the same channel, so you rely on clever in-flight compression rather than distributing work. Anthropic’s Claude takes a different route. It spins off sub-agents for each subtask, letting each agent tackle a piece of the problem in its own context window. Those mini-agents work in parallel, generate intermediate summaries via Haiku or Explore modules, then report back to a “parent” thread. You end up with faster visible output—several streams of tokens—but also extra cost. Sub-agents can duplicate effort, search the same files twice, or drop facts if they decide they aren’t worth passing up. On costs and coherence, the two camps diverge further. Claude’s multi-agent setup juggles messages more often, so you might pay for repeated searches or see inconsistencies when a sub-agent fails to forward a key detail. Codex’s compaction approach preserves more of the hidden state, reducing “forgetting” at the price of serial work. Anthropic’s model feels quicker because it spits out chunks in parallel, but that speed can mask missing context or repeated operations. Looking ahead, the author expects both firms to borrow from each other. OpenAI could layer in more agent-style delegation, while Anthropic tightens up its compression to avoid information loss. We’ll probably wind up with hybrids that mix continuous threads and targeted sub-agents, balancing coherence, speed, and cost.