Patch Notes B0.18 - GM Harness Improvements

You may hear us refer to the "GM harness" or "agent harness" from time to time. The "harness" refers to the code and tools that we give to the LLM, in addition to all of the instructions and context you provide. For example, the roll tool is something built into the GM's harness.

While world builders have a lot of control over the GM, there are some things that can only be done at the harness level, and some things where both the harness and the world builder can impact things. For example, token optimization is something both world builders and the harness can affect. World builders can build worlds and write instructions that do not consume too much context, and at the harness level, we can make sure that our tools don't return more context than they need to.

In this update, we make some significant updates to the GM harness.

• We improved the read/write/update accuracy by a lot. You should have much less instances of the GM attempting to write to a file 10+ times.
• We optimized the amount of context that is returned for each tool. We have stripped away meaningless tokens, like image URLs and other data that is not important to the LLM on reads. You should all see token savings, our estimate is about 10-20%. You may not see these token savings right away after the update as your historical context will not be changed, this is forward looking only.
• We massively improved the research tool. The tool is 10x faster now. We updated the research tool to use a special model that is extremely fast. It tends to add about 0.5-1 energy each time the tool is used. While that may seem expensive, the tool should save you a lot of context over subsequent turns. Instead of the GM reading 3 heavy pages, and those full pages living in your context window for following turns, the GM just gets a concise response back from the research tool. If you don't want to use the tool, you can turn it off in GM settings.
• We made additional optimizations that should increase your energy efficiency.
• We improved the system prompt by making it more clear and removed some confusing language.

Please continue playing and let us know if you feel a difference with anything. We don't expect there to be any massive differences in narration quality or anything, but we're curious if anyone has any positive or negative feedback as this was a pretty big code change.