An AI System Grows Upward, Not Sideways

An AI system has evolved beyond merely expanding sideways, expanding instead in an upward direction. For some time, adding more agents, workflows, and automation provided benefits, but eventually, the system reached a point where adding more of everything ceased to enhance performance. The next step did not come from broadening the scope, but rather from a higher layer.

The fleet had been operating prior to this overhaul, orchestrated by jobs scheduled on a timetable, pipelines written in TypeScript, miniature models performing preliminary checks, and sorting data. Each specialist was assigned to a single tool and nothing else. These initial arrangements were not temporary, but rather robust, reliable, and unremarkable, executing nightly, and they continue to do so today.

The initial design had inherent boundaries. The workers processed text and returned text, which was perfectly suited for tasks like sorting, compiling, and reporting. For any real intervention, such as altering code, assessing claims, or verifying assertions, a sequence had to be anticipated that accounted for every step. This type of system, once it reached this stage, did not grow further; it merely became more substantial.

Every new capability introduced another workflow, and each workflow required additional oversight for any subsequent changes.

However, the introduction of new tools altered the course. Agents reporting to a project could now operate autonomously, capable of opening files, modifying them, interacting with browsers, running tests, and interpreting their outputs. They connected to services and data sources through MCP, an open standard enabling programs to interface with external tools. Importantly, they initiated each task with access to a relevant portion of the system's memory. This integration occurred within Claude Code.

Rather than merely absorbing this change, a system with an existing fleet utilized it differently: it began to grow upward. Responsibilities shifted from the orchestrator to the supervisor. Instead of dictating the path for workers, the supervisor now issued assignments. An assignment is distinct from a planned sequence: it outlines a series of steps, the scope, rules, completion criteria, and stipulates when the agent should halt and request assistance.

An assignment is a comprehensive package, explicitly stating the goal, non-goals, scope, assigned files, permissible tools, and acceptance criteria.

This structured approach ensures that every assignment aligns with specified checks and stop conditions, reducing ambiguity and preventing undesirable outcomes. The supervisor's role shifts from planning to delegating, focusing on the assignment itself rather than the sequence of steps. This change is pivotal, as it allows for more efficient utilization of the system's capabilities and fosters an environment where agents can act independently, provided their actions are bound by the assignment's parameters.