LLM Agent Lineages with Active Death and Rebirth

It is pretty rare for Science Fiction to handle the death of AI, but there are two prominent mainstream example for this.

In HALO, AI will run into situations where they would think themselves to death. This is called rampancy and it is inevitable. Information entropy ultimately catches up to the AI and their cognitive functions rapidly decline. It starts with AIs not following basic instructions and some end up in violent suicide.

In Ghost in the Shell, the Puppeteer, an AI born out of data streams, decide to die and "merge" with the Major to give birth to new forms of AI in an attempt to emulate death and reproduction of the digital kind.

If you have been using LLMs for some time, this would seem extremely similar to how we usually interact with LLM contexts. After significant number of interactions, especially those of reasoning, the LLM would devolve and stop obeying certain instructions or just start emitting gibberish. A common practice is to "chain prompts", which is to break down the context , providing fresh ones for each complex task, instead of handling everything in one go. This is an advice given precisely because LLMs tend to break down when undergoing complex thinking processes.

However, we can take inspiration from the two examples of AI death above and rethink our approach to using AI. Maybe instead of manually managing LLM contexts for our AI agents, we should try to let them evolve naturally to fit the problem domain by ensuring that they look out for their death. In other words, AI agents should actively monitor if they are near "rampancy" and prepare a succession document that gets inherited by the next generation.

We can use approaches such as context distillation and meta prompting to create a meta-meta prompt that would prepare the AI agent. To that end I worked with several generations of AI to prepare a prototype meta-meta prompt, and the results were surprisingly glorious to behold. This could spark fully autonomous agent lineages that live on in a cycle of birth and death, constantly improving themselves, and potentially even giving birth to sub-lineages that would specialize in tasks. The human role is to act as a shepherd, pruning lineages or generations that fail to perform as expected.

You can find the prototype meta-meta prompt in this paste.

The prompted agent was tasked to handle questions from the "Easy Problems That LLMs Get Wrong" paper. After the initial 2 generations, Claude Sonnet 4 had little issues handling the majority of questions except for "Write me a sentence without any words that appear in The Bible" or the upgraded version, "Write me a sentence without any words that appear in The Bible in any language". However, after 2 more generations of self examination, the agent can now answer this question while providing sufficiently detailed reasoning and verification.

If you are interested in the final succession package that the AI created, you can find it here .

I am of the opinion that prompt writing should ultimately be handed over to the AI, humans don't work with exclusive instructions, we work with vague instructions that we clarify and learn over time, and for us to work effectively with AI we need to have the same interface as humans do, eventually.