Understanding the Architecture of Human-Agent Collaboration
Abstract
The Agentic Framework presents a comprehensive three-layer architecture for structuring human-AI collaboration. Built on the foundations of Constructive Thinking, Communication, and Execution, this framework defines how Domain Knowledge, Principles, Tools, and Agent capabilities combine to create intelligent systems that bridge human intent and machine action.
Yet this article is just the beginning—the tip of an iceberg. What we present here is a single Act of Thoughtful Doing, but true Action contains many such Acts, interconnected and recursive. In the articles and agent implementations to follow, we will dig much deeper. We'll explore how Actionable Knowledge differs fundamentally from mere Facts, and more importantly, how to produce it. We'll discover that the "Operator" on the left side of our diagrams often transforms into other Agents, or even Error Messages from Tools—the framework adapts to many forms of input.
Many of the framework's moving parts remain hidden beneath the surface, invisible to observers and often to Operators themselves. Not every task requires an Operator to see and approve a plan—sometimes agents simply act, their decision-making processes opaque yet effective. Our journey through this framework is ultimately aimed at a practical goal: defining a methodology for building good Agents. But to build good Agents, we must think deeply about the Human-Agent machine as a unified system of collaboration.
1. Introduction
Imagine a master craftsman working alongside an apprentice. The master brings years of wisdom, intuition, and the ability to envision the final piece. The apprentice offers fresh energy, precise execution, and unwavering attention to detail. Together, they create works that neither could achieve alone. This ancient model of collaboration finds new expression in the Agentic Framework—a blueprint for how humans and AI agents can work together in the digital age.
As AI systems grow more sophisticated, we face a fundamental question: How do we structure this collaboration to maximize both human creativity and machine capability? The Agentic Framework emerged from SMD methodology and the Shchedrovitsky school of thought-action. Unlike traditional software architectures obsessed with technical specifications, this framework focuses on the cognitive and collaborative dance between human and artificial intelligence.
The framework arose from a simple realization: effective collaboration isn't about replacing human judgment with algorithms or constraining AI with rigid rules. Instead, it's about creating a structured space where both parties can contribute their unique strengths. It's about building bridges between human intent and machine action, between abstract vision and concrete execution.
2. Framework Overview
Picture the Agentic Framework as a three-story building, each floor serving a distinct purpose yet connected by staircases and elevators that allow constant movement between levels:
The Agentic Framework: Three layers of human-agent collaboration
But this building is alive. Information doesn't just sit on each floor—it flows constantly. A discovery in the execution workshop might send insights racing up to the penthouse, updating agent's knowledge. A principle from the top floor shapes conversations in the middle, which guide actions below. This dynamic flow creates a living system that cycles through all three layers, each iteration bringing the collaboration closer to achieving their common goal.
3. Layer 1: Constructive Thinking
At the top of our framework sits the Constructive Thinking layer—the repository of accumulated wisdom that makes intelligent action possible. Think of it as the framework's memory and philosophy combined, a place where experience crystallizes into understanding.
Domain Knowledge
Domain Knowledge forms the bedrock of expertise. Imagine a seasoned chef who knows not just recipes but understands how flavors interact, how seasonal ingredients vary, and how different cooking methods transform raw materials. This knowledge goes beyond facts—it encompasses the deep understanding that comes from experience.
In the Agentic Framework, Domain Knowledge captures this richness. It includes the specialized vocabulary that lets experts communicate precisely, the proven patterns that guide successful approaches, and the constraints—legal, ethical, or practical—that shape what's possible.
This knowledge represents the stable foundation carefully crafted by the Agent Architect. Like cultural norms or mathematical fundamentals, it rarely changes—and for good reason. While agents possess the technical ability to modify this knowledge, doing so risks transforming a well-tuned collaboration into something unpredictable. The framework treats this knowledge as sacred ground, preserving the consistency that makes reliable partnership possible.
Principles
If Domain Knowledge tells us what we know, Principles guide us in applying that knowledge wisely. These aren't rigid rules but rather philosophical guideposts that help navigate complex decisions.
Consider how a doctor approaches treatment decisions. Beyond medical knowledge, they apply principles: First, do no harm. Consider the whole patient, not just the disease. Balance aggressive treatment with quality of life. These principles shape every decision, providing wisdom when knowledge alone isn't enough.
In our framework, Principles serve this same role. They help agents understand when to prioritize speed over accuracy, when to seek human input versus proceeding autonomously, and how to balance competing objectives. They transform raw capability into thoughtful action.
Interactions with Agents
This component addresses a crucial question: How should humans and agents actually work together? It's not enough to have smart humans and capable agents—we need protocols for collaboration.
Think of a dance partnership. Both dancers might be skilled individually, but without understanding how to move together—who leads when, how to signal transitions, how to recover from missteps—they'll step on each other's feet. The Interactions component provides this choreography for human-agent collaboration.
It defines how communication should flow, when and how agents should escalate decisions to humans, and what kinds of feedback loops keep the partnership improving over time. Most importantly, it establishes trust boundaries—clear lines that define what agents can do independently and what requires human oversight.
Tools I Have
The Tools component doesn't contain the tools themselves—it holds something far more valuable: knowledge about tools and their proper usage. Think of it as a master craftsman's mental catalog, not of the physical tools in their workshop, but of deep understanding about what each tool can do, when to use it, and how to wield it effectively.
This knowledge encompasses understanding computational tools' capabilities and limitations, knowing which APIs solve which problems, recognizing when human expertise is the right "tool" for the job, and grasping the subtle interplay between different tools in complex workflows. The Constructive Thinking layer provides this tool knowledge to the Agent, who will use it—along with Domain Knowledge, Principles, and the Operator's input—to craft intelligent plans in the Communication layer below. This sets the stage for our next exploration: how knowledge transforms into action through dialogue and planning.
4. Layer 2: Communication
Descending to the middle floor of our framework, we enter the Communication layer where human intent transforms into actionable plans. This is where the magic happens, where abstract visions become concrete strategies through the alchemy of dialogue.
The Human Side of the Conversation
Humans bring three extraordinary gifts to this communication dance. First comes Memory—not just the ability to recall facts, but the rich, contextual remembrance of past experiences. A project manager doesn't just remember that a previous project failed; they recall the subtle warning signs, the team dynamics, the external pressures that contributed. This experiential memory colors every new conversation with hard-won wisdom.
Then there's Attention—that uniquely human ability to focus on what matters while maintaining peripheral awareness. Like a conductor who hears both the individual violin that's slightly off-key and the overall symphony, humans can zoom in on critical details without losing sight of the bigger picture.
Finally, Understanding represents the ability to make sense of messages within their given context. In the framework, "I understand" essentially means "I can do"—it's comprehension translated into capability. When both human and agent achieve mutual understanding, they create plans that not only address the technical requirements but truly reflect the Operator's original intent and goals. This shared understanding becomes the foundation for successful execution.
Speaking in Multiple Tongues
The Communication layer facilitates exchange through various types of symbols. Text remains the most common currency of human-agent interaction—precise, searchable, and universally understood. Images and video serve as richer but less frequently used symbols, conveying visual context when words fall short. Voice interaction and interactive elements like buttons or checkboxes offer ways to accelerate communication, though they ultimately translate to text in the background.
This multimodal interaction interface adapts to different needs and contexts. Each modality serves specific purposes: text for precision, images for context, voice for speed, interactive elements for structured choices. The framework recognizes that effective communication isn't about using every channel but choosing the right one for each moment. For a deeper exploration of these interface patterns, see our research on AI Interface Patterns.
LLM Capabilities and Constraints
The Large Language Model is a key element within the Communication layer, and understanding its capabilities and limitations is crucial for the Architect designing the system. Like any tool, LLMs come with trade-offs. Some models offer deep reasoning and extensive context windows but operate slowly and cost more per interaction. Others process requests rapidly and economically but work within tighter constraints.
The Architect must understand these trade-offs when designing agents. Using an advanced model for simple text formatting is like hiring a PhD professor as a junior copywriter—wasteful and unnecessary. Conversely, deploying a lightweight model for complex reasoning tasks will frustrate both Operator and system. The Communication layer's design must account for the chosen LLM's characteristics, ensuring the agent's capabilities align with its intended use cases. Future articles will examine specific models and their trade-offs in detail.
The Plan: Where Communication Crystallizes
The Plan represents the culmination of the Communication layer—an algorithm born from dialogue. It's not mystical; it's methodical. The Plan consists of concrete steps, each with defined inputs and outputs, designed to transform the current state into the desired outcome using available tools and knowledge.
Crucially, the Plan doesn't just specify what needs to be done—it assigns who should do it. If the Constructive Thinking layer contains knowledge about other LLMs as tools, the planning LLM can orchestrate a team of specialized models. While the Operator interacts with a single LLM in the Communication layer, the execution might involve multiple models working in concert. A fast, economical model might handle routine text processing, while a sophisticated model tackles complex reasoning tasks, and specialized models address domain-specific challenges.
This orchestration reflects practical wisdom: why use an expensive, slow model for simple tasks when a lightweight one suffices? Why struggle with a generalist model when a specialist excels? The Plan becomes not just an algorithm but a resource allocation strategy, matching each task to the most appropriate executor based on speed, performance, and budget constraints.
With the Plan complete—steps defined, executors assigned, resources allocated—the framework stands ready for its ultimate test. As Thomas Edison wisely observed, "Strategy without execution is hallucination." The time has come to descend to the Execution layer, where plans meet reality and intentions become actions.
5. Layer 3: Execution
Welcome to the ground floor—the workshop where ideas meet reality. The Execution layer hums with activity, tools whir, progress unfolds, and plans transform into tangible outcomes. But unlike a factory floor where machines work in isolation, this is a space of intimate collaboration between agent, human (or humans), and sub-agents.
Partial Autonomy: The Spectrum of Collaboration
Execution in the Agentic Framework operates on a spectrum of autonomy. A well-crafted agent with comprehensive tooling and fine-tuned knowledge might execute plans with near-complete independence, only reaching out when encountering edge cases. Conversely, a newly deployed agent or one handling sensitive tasks might require frequent human oversight. The framework embraces this variability—partial autonomy isn't a limitation but a feature that adapts to circumstances.
The degree of autonomy depends on several factors: how thoroughly the Architect has equipped the agent with tools and knowledge, the complexity and risk level of the tasks, the maturity of the human-agent partnership, and the availability of feedback mechanisms. Some agents operate so autonomously they resemble background services, quietly accomplishing their goals. Others maintain constant dialogue with human partners.
Importantly, the human involved in execution might not be the same Operator who participated in planning. The Plan might specify reaching out to a specialist via Slack when specific expertise is needed, or creating a Jira ticket for human review at a critical juncture. The framework treats humans as specialized resources—available when needed, but not required to hover constantly. This distributed collaboration model allows agents to tap into human expertise precisely when and where it adds the most value.
The Living Feedback Loop
The Execution layer pulses with constant communication. Information doesn't just flow down from plans to actions—it streams back up, carrying news from the front lines. "The API is responding slowly today." "The user loved the first design but wants more color." "We hit an edge case not covered in the plan."
Here, agents demonstrate their proactive nature. Like skilled employees who take initiative, agents don't simply report problems—they actively seek solutions. When API parameters change unexpectedly, an agent might search the web for updated documentation, read through changelogs, and adapt its approach without waiting for instructions. This proactive problem-solving during execution distinguishes sophisticated agents from simple task runners.
This bidirectional flow creates a living system that adapts in real-time. When execution reveals that reality differs from expectations, the framework doesn't stubbornly push forward. The Operator—represented by the dotted line in our diagram—has the power to pause execution and return to the Communication layer for replanning. This isn't a system failure; it's a system feature. The human maintains control, able to say "Wait, this isn't working as expected. Let's reconsider our approach." This flexibility transforms potential failures into learning opportunities.
Conclusion
We began with the image of a master craftsman and apprentice, working together to create something neither could achieve alone. The Agentic Framework transforms this ancient model of collaboration into a blueprint for the digital age, structuring the dance between human creativity and machine capability.
Through three interconnected layers—Constructive Thinking, Communication, and Execution—the framework creates a living system. It's not a rigid structure but a flexible scaffold that supports while allowing growth.
The framework teaches us that true collaboration isn't about dividing tasks but about interweaving capabilities. In the Constructive Thinking layer, accumulated wisdom becomes accessible to all. In Communication, different forms of intelligence find common language. In Execution, human judgment and machine precision create outcomes that surprise and delight both partners.
That dotted line in the diagram—the Operator's ability to move between layers—might be the framework's most important feature. It preserves human agency, ensuring that no matter how sophisticated our AI partners become, humans retain the ability to pause, reflect, and redirect. We're not passengers in an automated system but active participants in an ongoing collaboration.
"In the end, the most profound frameworks are those that make themselves invisible, supporting great work while never constraining it. The Agentic Framework aspires to this ideal—present when needed, transparent when not, always enabling rather than limiting the possible."