The increasing sophistication of AI agents presents a critical question for users: when is a single, capable agent sufficient, and when does the complexity of an agent team become necessary for enhanced quality and efficiency? This exploration delves into the tiered approach of agentic workflows, from single agents with sub-agent delegation to intricate Kanban graphs, examining the requirements and optimal use cases for each level.

The Foundation: Single Agent with Sub-Agent Delegation

At its core, an agentic workflow can begin with a single, powerful agent capable of delegating tasks to temporary "sub-agents." This approach allows a primary agent to manage a complex problem by breaking it down into smaller, manageable components. The parent agent initiates the process, defines the sub-tasks, and then synthesizes the results.

A key benefit of this delegation model is token efficiency. Instead of the primary agent processing all information and potentially exceeding its context window, it can offload specific research or processing tasks to sub-agents. These sub-agents, operating with their own context, return summarized information to the parent agent. This strategy conserves the primary agent's context window, allowing it to focus on higher-level judgment and synthesis. For instance, a research task that might consume 102,000 tokens for a single agent could be reduced to 37,000 tokens when delegated to sub-agents, with the parent agent receiving compressed, synthesized reports.

Crucially, sub-agents begin with a fresh conversational context, possessing no prior knowledge of the parent agent's history or tool calls. Their sole context is derived from the goal and specific instructions provided by the parent agent during the delegation process. This ensures that each sub-task is executed within its defined parameters.

Enhancing Quality with Goal Loops

To further refine the output of a single agent or a delegated workflow, a "goal loop" can be implemented. This feature allows users to define a standing objective that the agent pursues across multiple turns until completion. A goal acts as a meta-prompt, linking a series of individual prompts into a broader task.

The agent is instructed on what constitutes a successful outcome, how to verify it, and when to seek human intervention. The main agent then functions as a supervisor, monitoring progress against the defined goal. For example, a research task can be augmented with specific criteria, such as requiring a minimum number of peer-reviewed sources and a word count limit for the final report. The agent will then iteratively work towards meeting these criteria, potentially looping through sub-agent delegations and cross-referencing information until the goal is achieved. This iterative process can be extended, allowing for continuous refinement and quality assurance, even for extended periods, such as an agent working on a coding project for hours, automatically testing and repairing code.

Care must be taken with resource allocation when using goal loops, especially with advanced models. Specifying lower-intelligence models for sub-tasks can significantly conserve computational resources and costs, reserving the higher-intelligence models for critical decision-making and synthesis. This can be configured through agent profiles or specific model settings for auxiliary tasks.

Introducing Specialization: Bot Mode and Group Chats

As workflows become more complex, the need for specialized agents with persistent knowledge and skills arises. This is where "bot mode" becomes valuable. Bot mode allows for the creation of distinct agent profiles, each equipped with its own skills, memory, and working context, potentially utilizing different models suited to their roles.

These specialized agents, or "bots," can be configured with specific default models. For example, an "orchestrator" bot might use a frontier model, while a "researcher" bot could utilize a less powerful, more cost-effective model. These bots can then collaborate through group chats, forming a communication graph. Unlike sub-agents, these specialized bots retain their context and memory across conversations, allowing for more sophisticated and persistent workflows.

A typical group chat scenario might involve an orchestrator bot assigning tasks to a researcher bot for investigation and a librarian bot for knowledge base management. The orchestrator acts as the human proxy, directing the flow of information and tasks between specialized agents. While group chats offer a form of graph-based interaction, they are limited in conversational turns before requiring human intervention.

Managing Complex Workflows: Kanban Graphs

For highly complex, multi-step processes, a Kanban graph, or "Conbon" system, offers a robust solution. Conbon provides a durable alternative to simple delegation, featuring named profiles, shared state, dependencies, retries, and work that persists beyond a single session.

Key distinctions between Conbon and simple delegation include:

• Orchestrator Blocking: In delegate task, the parent agent is blocked until sub-agents complete their work. Conbon allows the orchestrator to assign tasks and continue, with results returning only when explicitly configured.
• Persistent Memory: Conbon utilizes named profiles with persistent memory, unlike anonymous sub-agents that start with a fresh context.
• Resumability: Conbon tasks can be blocked and resumed from their point of failure, offering greater durability than sub-agents, which simply report completion status.
• Human-in-the-Loop: Conbon facilitates explicit human intervention at specific task stages, whereas delegation offers limited steering capabilities.
• Audit Trail: Conbon stores task history in an SQLite database, providing a permanent, transparent audit trail, unlike delegation where context can be lost due to compression.

Conbon operates as a work queue, where each task handoff is a database entry, visible and editable by any profile or human user. This transparency forms a control graph, illustrating the dependencies and flow of tasks.

Within Conbon, tasks progress through various stages:

• Triage: Raw ideas are refined into actionable tasks by a "specifier" agent.
• To-Do: Tasks awaiting dependencies or assignment.
• Scheduled: Tasks set to run at a specific time.
• Ready: Dependencies are met, and the task is assigned.
• Running: A worker agent is actively processing the task.
• Blocked: Waiting for human input.
• Review: A dedicated review agent assesses the work.
• Completed: The task is finished.

The Conbon system can be configured with specific auxiliary models for triage and decomposition. A powerful model can be used for decomposing complex tasks into smaller, manageable subtasks, while less powerful, more cost-effective models can handle the execution of these individual tasks. This strategic allocation of resources can significantly reduce operational costs.

Projects can be organized using boards and projects, allowing for distinct task graphs and configurations tailored to specific workflows, such as administrative, research, or coding projects. The orchestration settings allow for the designation of an orchestrator profile, which manages task assignment, and a default assignee.

When creating a task, users can set priorities, specify workspaces, list skills, and choose models. Crucially, tasks can be run in "goal mode," enabling individual worker agents to operate within their own self-contained goal loops. Cost estimation is also available, providing an approximation of token usage for a given task, allowing for adjustments to manage budget constraints.

Conclusion: Strategic Implementation of Agentic Workflows

The choice between a single agent with delegation, goal loops, bot mode, or a full Conbon system depends on the complexity and specific requirements of the task. For simple tasks, a single agent with delegation may suffice. As complexity increases, introducing goal loops for iterative refinement, bot mode for specialized persistent agents, and Conbon for robust workflow management becomes beneficial.

The overarching principle is to start with the simplest level of coordination that meets the immediate needs and to incrementally add complexity only when a specific problem necessitates it. This strategic approach ensures that resources are utilized efficiently and that the chosen workflow effectively addresses the desired outcomes, transforming incoming information into actionable insights and reusable knowledge.