OpenAI and Cursor both embrace multi-agent architecture, but split on who coordinates

OpenAI's recently announced Agents API and Cursor's new Projects feature arrived within hours of each other in September, both exposing a similar architectural pattern: a coordinator breaks complex software tasks into smaller work units assigned to specialized agents, rather than forcing a single agent to manage everything at once. While the platforms operate at different layers, they converge on the same principle—dividing labor improves reliability and accuracy.

This pattern has precedent. AWS Bedrock AgentCore reached general availability in October 2025, and Anthropic rolled out Claude Managed Agents in April 2026. What distinguishes the OpenAI and Cursor announcements is that two heavyweight competitors in AI-assisted development arrived at the same architectural solution independently and simultaneously, signaling industry-wide recognition that the single-agent model has limits.

Why single-agent workflows break down#

A conventional coding agent observes a repository, reasons about next steps, invokes tools, and examines results in a loop. For small tasks this works fine, but complex work—such as migrating a database, updating dependent services, rewriting tests, and validating the system—stretches that loop to its breaking point.

The core problem is context degradation. Maintaining all relevant information across multiple steps within a single conversation window becomes increasingly unreliable as the context grows. Researchers have documented this phenomenon as context rot: a 2026 study of frontier models including Claude Opus, GPT-5.4, and Gemini 3.1 Pro found they missed critical information buried in transcripts after 800,000 tokens of routine activity—two to 30 times more often than they caught it early on.

Additionally, forcing sequential execution of independent tasks converts what should be parallel work into a bottleneck. A single agent cannot efficiently analyze a database schema, review documentation, and discover test gaps all at once.

The coordinator-subagent split#

The coordinator-subagent model inverts this approach. Rather than asking one agent to juggle the entire context, the coordinator acts as a control plane: it understands the overarching objective, breaks work into discrete units, and assigns each to a specialized agent operating in its own isolated context. GitHub's custom-agent framework exemplifies this, giving each agent only the prompts, tools, and environment it needs.