Artifacts can reduce token usage by a factor of 2 or more.
I had an [agent evolution framework](danieltan.weblog.lol/2025/06/agent-lineage-evolution-a-novel-framework-for-managing-llm-agent-degradation) before that dumped the output analysis into chat. It often timed out before the 10th conversation. After dumping the analysis into an artifacts, and have the LLM only edit it as required, I can go to 15 or more rounds without hitting the context limit. While they seem to re-output the entire artifact each time, they don't actually consume the tokens for the entire artifact.
This also greatly reduces the tendency of HALO-style rampancy, or AI psychosis which is also what the recent paper on context-rot/poisoning (https://research.trychroma.com/context-rot) is about.
I had an [agent evolution framework](danieltan.weblog.lol/2025/06/agent-lineage-evolution-a-novel-framework-for-managing-llm-agent-degradation) before that dumped the output analysis into chat. It often timed out before the 10th conversation. After dumping the analysis into an artifacts, and have the LLM only edit it as required, I can go to 15 or more rounds without hitting the context limit. While they seem to re-output the entire artifact each time, they don't actually consume the tokens for the entire artifact.
This also greatly reduces the tendency of HALO-style rampancy, or AI psychosis which is also what the recent paper on context-rot/poisoning (https://research.trychroma.com/context-rot) is about.