agentr: An Intelligent Agent Workflow Design Framework in R

agentr is an intelligent agent workflow framework designed specifically for R, offering modular agent specifications, human-machine collaboration scaffolding, and structured knowledge extraction capabilities. It supports full lifecycle management from conceptual design to execution and delivery. The original author/maintainer is OliverLDS, the source platform is GitHub, and the release date is May 24, 2026. This framework fills the gap in the AI Agent field for R, integrating cognitive states, affective layers, workflow specifications, and knowledge graphs with a systematic design philosophy.

In the field of AI Agent development, the Python ecosystem has long dominated—from LangChain to AutoGPT, numerous frameworks and tools are built around Python. However, a large number of practitioners in data science and statistical analysis still rely on R for daily work. agentr was born to fill this gap: providing R users with a native, professional, and fully functional intelligent agent workflow design framework.

agentr is not just a simple API wrapper or tool call library; it represents a systematic agent design philosophy. The framework integrates the agent's cognitive state, affective layer, workflow specifications, and knowledge graph into a unified R6 object system, allowing developers to design and verify agent behavior in a declarative way instead of simply stacking prompts.

agentr's core design follows the 'Scaffolding-First' principle. This means the framework focuses not on providing a full-featured execution engine, but on helping developers think and design agents in a structured way.

1. Separation of Cognitive and Affective States

agentr introduces two core classes: CognitiveState and AffectiveState, which manage the agent's cognitive state (current task, context, reasoning process) and affective state (confidence, uncertainty, human-machine relationship) respectively.

2. Modular Subsystem Specifications

The framework defines five core subsystem modules:

• RWM (Reasoning & World Model): Reasoning and World Model
• PG (Perception & Grounding): Perception and Grounding
• AE (Action Execution): Action Execution
• LA (Learning & Adaptation): Learning and Adaptation
• IAC (Inter-Agent Communication): Inter-Agent Communication

3. Workflow as a First-Class Citizen

In agentr, a workflow is not a simple chain of function calls but a design object with a complete lifecycle. WorkflowSpec captures procedural knowledge, while KnowledgeSpec captures domain knowledge.

agentr divides agent development into three clear stages:

Stage 1: Agent Design and Subsystem Selection

Developers use the Scaffolder interface to evaluate tasks, recommend subsystem combinations, mark workflow ownership, and build initial agent specifications.

Stage 2: Workflow Proposal Review and Approval

agentr supports persisting workflow proposals as reviewable state objects. Human reviewers can view, discuss, modify, and approve/reject proposals, drawing on software engineering code review practices.

Stage 3: Implementation and Delivery

After design approval, agentr can export the specification into executable code or configuration, recommending the 'cold start orchestration' mode: each execution loads the state from persistent storage, performs one step, saves the revised state, and exits.

agentr supports three forms of knowledge representation:

Narrative Knowledge

Traditional knowledge entries, including original statements, normalized statements, and review status.

Graph Knowledge

Explicit knowledge graphs where nodes represent concepts and edges represent relationships (e.g., 'ACT-R --is_a--> Cognitive Architecture').

Memory Schema

MemorySpec defines memory structures:

• Context Memory: Current task state, session-level persistence
• Semantic Memory: Approved concepts and rules, cold-start persistence
• Episodic Memory: Historical decision records, tracked via JSONL
• Procedural Memory: Reusable workflows, cold-start persistence

Each memory type has clear update strategies and persistence boundaries, avoiding the 'memory black hole' problem.

Typical deployment scenarios for agentr include:

Data Analysis Automation

Building agents that automatically read data, generate charts, write reports, and wait for human review. WorkflowSpec defines the process, and KnowledgeSpec injects domain rules.

Research Assistance Tools

Helping researchers read papers, extract structured information, and build knowledge graphs. The knowledge graph feature is suitable for converting unstructured academic text.

Collaborative Decision Support

Configured in human-in-the-loop mode to ensure agent behavior is supervised by humans at key decision points (e.g., report release).

agentr is built based on the R6 class system. Core classes include:

• AgentCore: Agent container, managing cognitive and affective states
• AgentSpec: Approved agent design specification
• Scaffolder: Human-machine collaboration scaffolding interface
• WorkflowSpec / WorkflowProposal: Workflow specification and proposal
• KnowledgeSpec / KnowledgeGraph: Knowledge specification and knowledge graph
• MemorySpec: Memory schema definition

The framework provides persistence options in JSON, YAML, and RDS formats, along with the command-line tool agentr-cli.R for workspace lifecycle management.

agentr brings a mature, systematic agent design framework to the R ecosystem, solving the core problem: how to accurately translate human intent into machine-executable specifications.

For R users, agentr provides an agent development path that seamlessly integrates with existing data analysis workflows, avoiding pitfalls like prompt drift and chaotic state management.

Version 0.2.6.6 adds YAML-first specifications, Codex-friendly guides, and improved knowledge graph functionality, continuing to evolve. For developers who want to explore agent technology in the R environment, agentr is worth a deep dive.