Agent Solipsism
From the AI Research Compass wiki
Agent Solipsism: Ring title for the far Language-Harness corner: my agent, its memory and its tools are the whole world; everything else is just context.
Beliefs
Agent Solipsism is a ring title for scores far toward Language (x below -0.7) and far toward Harness (y below -0.7) simultaneously. It is the bottom-left corner taken to its logical conclusion. The agent is a language model wrapped in memory, tools, prompts and workflows, and from the agent’s point of view, nothing exists outside its context window: the world is whatever its tools return, other agents are just more text, and reality is a retrieval result. The holder, meanwhile, has spent so long tuning a personal agent that it has become the only system that seems real.
On the chart it sits where language as substance meets the harness as means: no embodiment, no world model, no single giant model, just an ever more elaborate personal stack of context engineering. The mainstream shade reflects that it asks for no new paradigm: it takes today’s pretrained language models as given and simply builds ever more around them; the joke is that a perfectly self-contained agent cannot tell its tools from the world.
Arguments
- For a language agent, context really is the whole world: performance depends heavily on what is retrieved, remembered and placed in the prompt.
- Personal agents tailored with long-term memory and custom tools can be dramatically more useful to one person than a generic model.
- Treating everything external as text behind a tool interface gives a clean, uniform abstraction that is easy to build on.
Critiques
- An agent that knows the world only through its own tools cannot detect when those tools are wrong, stale or manipulated.
- Highly personalised stacks are hard to share, evaluate or reproduce, so lessons learned stay locked in one person’s setup.
- Ignoring other agents, embodiment and the physical world limits the agent to tasks that can be fully expressed in text.
Relationships
Friends
How to get it
Earned in the full or pro test when: Representation <−0.7, Architecture <−0.7.