Infrastructure · Data Structures
Context
Portable execution environment with inherited constraints
Definition
Mechanism
The portable execution environment and state container for an agent. It encapsulates: (1) Inherited Constraints (Safety/Budget), (2) Available Tools/Capabilities, (3) Working Memory (History/Variables), and (4) Identity Claims. Context flows through delegation chains, acting as the 'Stack Frame' of the agentic system. It allows agents to be paused, moved, or cloned.
Preconditions
- Parent context exists (or Root is created)
- Schema is defined
Postconditions
- Agent has a valid environment to execute.
Invariants
- Constraint Monotonicity: Child contexts can ADD constraints but never REMOVE inherited ones.
- Serializability: The entire context must be serializable to JSON/Storage for persistence.
- Provenance: Every item in the context must be traceable to its source (Observation, User Input, or Inference).
Risks
Failure modes
Context Contamination/Leak
Private data from one context leaking into another (e.g., across tenant boundaries).
Context Poisoning
malicious prompts or data injected into the context to hijack agent behavior.
Stale Context
Acting on outdated information after a long pause.
Context Explosion
Size grows beyond the window limit of the model.
Relationships
Builds on
References
Non-negotiable boundary condition
Unique agent context (The Who)
Used by
References
Root primitive for state modification
Society
Isolate agent execution with resource limits and egress controls
Mind
Structural limit on parallel execution (Fan-out)
Mind
Semantic compression for long-running contexts
Mind
Load state before acting
Society
Ingest only what is strictly necessary
Mind
Postponing decisions until more information available
Society
Explicit safety reasoning prior to execution
Mind
Predictive modeling of external agent states
Society
Generate-evaluate-refine loop
Mind
Local action via global context embedding
Mind
Name the larger situations a moment sits inside, and what each one constrains
Mind
Breaking local optima
Mind
Simulating diverse experts
Mind
Self-location awareness within the task graph
Mind
Clarify intent and constraints before planning
Mind
Ground responses in retrieved external knowledge
Mind
Reinforce constraints in system prompt before each response
Mind
Navigate like a researcher, not linearly
The atomic unit of intent
Execute external tool and observe result
Society
Claim tasks and become the assigned persona
Accepts
Society
Explicit, scoped protocol-mode toggle with enter/revert boundaries
Mind
Concept blending with novelty and value acceptance
Society
Transfer control and context between agents
Mind
Execute a cognitive topology
Mind
Atomic cognitive step
Yields
Active State Perception
Composes with
Mind
Rapid decision cycle: Observe, Orient, Decide, Act
Reference
Data schema
{
"type": "object",
"properties": {
"id": {
"type": "string"
},
"constraints": {
"type": "array"
},
"history": {
"type": "array"
},
"variables": {
"type": "object"
}
}
}