Instructions API Reference

This page contains the auto-generated API documentation for the ArbiterOS instruction types.

The instruction types define the fundamental operations that can be governed by ArbiterOS. Each instruction belongs to one of eight core categories, representing different aspects of agent behavior.

Instruction Type Union

InstructionType = Union[CognitiveCore, MemoryCore, ExecutionCore, NormativeCore, MetacognitiveCore, AdaptiveCore, SocialCore, AffectiveCore] module-attribute

CognitiveCore

Governs probabilistic reasoning. Its outputs are always treated as unverified until subjected to explicit checks.

CognitiveCore

Bases: Enum

Governs probabilistic reasoning. Its outputs are always treated as unverified until subjected to explicit checks.

GENERATE = auto() class-attribute instance-attribute

Invokes the LLM for text generation, reasoning, or formulating a query. This is the most general-purpose cognitive instruction, producing content including text, hypotheses, and queries.

DECOMPOSE = auto() class-attribute instance-attribute

Breaks a complex task into a sequence of smaller, manageable sub-tasks or creates a formal plan of execution. Transforms complex problems into structured, actionable components.

REFLECT = auto() class-attribute instance-attribute

Performs self-critique on generated output to identify flaws, biases, and areas for improvement. Often produces a structured critique to guide subsequent GENERATE steps and self-diagnosis of prior outputs.

MemoryCore

Manages the LLM's limited context window and connections to persistent memory.

MemoryCore

Bases: Enum

Manages the LLM's limited context window and connections to persistent memory.

LOAD = auto() class-attribute instance-attribute

Retrieves information from an external knowledge base (e.g., a vector store or document) to ground the agent.

STORE = auto() class-attribute instance-attribute

Writes or updates information in long-term memory, enabling agent learning and persistence.

COMPRESS = auto() class-attribute instance-attribute

Reduces the token count of context using methods like summarization or keyword extraction to manage the limited context window.

FILTER = auto() class-attribute instance-attribute

Selectively prunes the context to keep only the most relevant information for the current task.

STRUCTURE = auto() class-attribute instance-attribute

Transforms unstructured text into a structured format (e.g., JSON) according to a predefined schema.

RENDER = auto() class-attribute instance-attribute

Transforms a structured data object (e.g., JSON) into coherent natural language for presentation to a user.

ExecutionCore

Interfaces with deterministic external systems. These are high-stakes actions requiring strict controls.

ExecutionCore

Bases: Enum

Interfaces with deterministic external systems. These are high-stakes actions requiring strict controls.

TOOL_CALL = auto() class-attribute instance-attribute

Executes a predefined, external, deterministic function (e.g., API calls, database queries, code interpreters). Provides structured interaction with vetted external services.

TOOL_BUILD = auto() class-attribute instance-attribute

Writes new code to create novel tools on-the-fly. Enables dynamic capability extension through programmatic generation of custom functions and utilities.

DELEGATE = auto() class-attribute instance-attribute

Passes sub-tasks to specialized agents in multi-agent systems. Facilitates hierarchical task decomposition and leverages domain-specific expertise across agent networks.

RESPOND = auto() class-attribute instance-attribute

Yields final, user-facing output and signals task completion. Serves as the terminal instruction in any execution workflow, ensuring proper task closure.

NormativeCore

Enforces human-defined rules, checks, and fallback strategies. This domain anchors ARBITEROS's claim to systematic reliability.

NormativeCore

Bases: Enum

Enforces human-defined rules, checks, and fallback strategies. This domain anchors ARBITEROS's claim to systematic reliability.

VERIFY = auto() class-attribute instance-attribute

Performs objective correctness checks against verifiable sources of truth (e.g., schemas, unit tests, databases).

CONSTRAIN = auto() class-attribute instance-attribute

Applies normative compliance rules ('constitution') to outputs, checking for safety, style, or ethical violations.

FALLBACK = auto() class-attribute instance-attribute

Executes predefined recovery strategies when preceding instructions fail (e.g., failed TOOL CALL).

INTERRUPT = auto() class-attribute instance-attribute

Pauses execution to request human input, preserving agent state for oversight.

MetacognitiveCore

Enables heuristic self-assessment and resource tracking, supporting adaptive routing in the Arbiter Loop.

MetacognitiveCore

Bases: Enum

Enables heuristic self-assessment and resource tracking, supporting adaptive routing in the Arbiter Loop.

PREDICT_SUCCESS = auto() class-attribute instance-attribute

Estimates the probability of successfully completing the current task or plan, providing anticipatory assessment of feasibility.

EVALUATE_PROGRESS = auto() class-attribute instance-attribute

Performs strategic assessment of the agent's current reasoning path to answer heuristic, goal-oriented questions about viability and productivity.

MONITOR_RESOURCES = auto() class-attribute instance-attribute

Tracks key performance indicators including token usage, computational cost, and latency against predefined budgets and thresholds.

AdaptiveCore

Governing autonomous learning and self-improvement within the ArbiterOS paradigm.

AdaptiveCore

Bases: Enum

Governing autonomous learning and self-improvement within the ArbiterOS paradigm

UPDATE_KNOWLEDGE = auto() class-attribute instance-attribute

Integrates new information into knowledge base via autonomous curriculum generation, web data retrieval, and distillation processes

REFINE_SKILL = auto() class-attribute instance-attribute

Improves existing capabilities through self-generated code testing, fine-tuning on new data, or techniques like Self-Taught Optimizer (STOP)

LEARN_PREFERENCE = auto() class-attribute instance-attribute

Internalizes feedback from human interaction or environmental rewards via Direct Preference Optimization (DPO) or Reinforcement Learning from Human Feedback (RLHF)

FORMULATE_EXPERIMENT = auto() class-attribute instance-attribute

Designs and proposes experiments for active learning loops to discover environmental properties or self-capabilities

SocialCore

Enabling governable inter-agent collaboration in multi-agent systems.

SocialCore

Bases: Enum

Enabling governable inter-agent collaboration in multi-agent systems

COMMUNICATE = auto() class-attribute instance-attribute

Sends a structured message to another agent, following a defined protocol for inter-agent coordination

NEGOTIATE = auto() class-attribute instance-attribute

Engages in a multi-turn dialogue with another agent to reach a mutually acceptable agreement on a resource or plan

PROPOSE_VOTE = auto() class-attribute instance-attribute

Submits a formal proposal to a group of agents and initiates a consensus-forming protocol

FORM_COALITION = auto() class-attribute instance-attribute

Dynamically forms a temporary group or 'crew' of agents to tackle a specific sub-task, defining roles and shared objectives, as seen in frameworks like CrewAI

AffectiveCore

Enabling governed socio-emotional reasoning for human-agent teaming.

AffectiveCore

Bases: Enum

Enabling governed socio-emotional reasoning for human-agent teaming

INFER_INTENT = auto() class-attribute instance-attribute

Analyzes user communication to infer underlying goals, preferences, or values that may not be explicitly stated

MODEL_USER_STATE = auto() class-attribute instance-attribute

Constructs or updates a model of the user's current cognitive or emotional state (e.g., confused, frustrated) based on interaction history

ADAPT_RESPONSE = auto() class-attribute instance-attribute

Modifies a planned response to align with the user's inferred state or established preferences (e.g., adjusting tone, verbosity, or level of detail)

MANAGE_TRUST = auto() class-attribute instance-attribute

Evaluates the history of interactions to assess the level of trust the user has in the agent and proposes actions to build or repair that trust

Short Aliases

For convenience, all instruction types are available as module-level constants:

import arbiteros_alpha.instructions as Instr

# Use short aliases instead of full paths
@os.instruction(Instr.GENERATE)  # instead of Instr.CognitiveCore.GENERATE
def generate(state): ...

@os.instruction(Instr.TOOL_CALL)  # instead of Instr.ExecutionCore.TOOL_CALL
def tool_call(state): ...