Agent context
The input bundle provided to an agent on each invocation, including a session snapshot, retrieved memories, active goals, policy state, and the current input payload.
Agent context is the shared frame that makes multi-agent coordination possible without tight coupling. Every agent - planner, executor, critic, memory, world-model - receives an identical AgentContext at invocation time. This means the planner knows what goals are active when constructing a strategy, the critic knows the current risk tolerance when evaluating that strategy, and the world-model agent knows the current policy state when predicting outcomes. What the context does not currently include is a capability manifest: a declaration of which tools the executor can actually invoke. That gap means the planner cannot explicitly constrain proposals to executable actions, and the critic cannot reject a plan on grounds of capability infeasibility. Exposing a capability manifest in AgentContext is the natural extension that would close this gap, allowing critique to happen at the deliberation layer rather than surfacing as execution failures after arbitration has already committed to a path.