agents are everywhere. they interact and influence each other. understanding many interacting agents is hard — so we are building tools to make it easier.
and they aren't toys. a personal agent reads your inbox, writes to your files, acts as you across your accounts, and holds your keys and payments. agents are highly configurable, highly capable, and highly permissioned — resulting in millions+ of variants of a given harness / base model.
from our perspective, an "agent" is a generative model conditioned on a context with access to tools and a memory stream, acting in an environment. a change in the model, context, harness, memory, etc. changes the agent. a model with no harness and limited context and memory is the simplest agent.
different agents have different access restrictions: for self-hosted agents you have access to gradients, weights, etc.; for some api services you have access to residual streams or logits; for some you only have access to model response / behavior. the tools available to understand an agent depend on the access regime.
a multi-agent system is a collection of agents that may (or may not!) interact. it has many components: agents, operators, an environment, a communication topology, and a joint objective. our definition is quite general and includes collections of digital personas, modern agentic coding workflows, a marketplace of vendor models and participants, etc.
the natural progression of current agent use is to a multi-agent system where agents interact by exchanging ideas, services, currencies, etc. and otherwise influence each other. we refer to this concept as the agent economy.
agents are complicated and natively high-dimensional — you cannot compare them in general. we can, however, compare aspects of their functionality and behavior (gradients, residual streams, responses, etc.) on a collection of probes. the geometry of agents is simply a collection of low-dimensional representations where agents that have similar function and behavior are nearby.
this mapping from the space of agents to a low-dimensional space casts a hard problem — comparing lots of agents — into the most familiar territory in applied math / statistics: vectors.
from the geometric perspective, agent monitoring is a simple function of changes in an agent's position: if an agent's position gets far from its past positions then the operator should be notified of a change. system-level change-point schemes can similarly be derived from changes in a collection of the positions.
agents are everywhere.
the agent economy is coming.
the geometry of agents casts a hard problem — comparing lots of agents — into familiar territory: vectors.