Agent Sprawl Becomes the Real Problem

The agent stack is growing up. Shopify's public-Slack experiment, OpenHands' new control plane, and recursive goal patterns all point the same direction: orchestration, not intelligence, is the bottleneck.

The agent conversation this week shifted from "what can it do?" to "how do we manage hundreds of them?" Three threads converged: Shopify making their internal agents public-by-default in Slack, OpenHands shipping a control plane for enterprise sprawl, and developers like Jason Liu nesting goals inside goals. The orchestration problem is now bigger than the model problem.

The most telling design choice this week came from Shopify, where Simon Willison noted that their River agent system only operates in public Slack channels — explicitly so other employees can learn by watching. Willison compared it to Midjourney's Discord-only launch, where the weird craft of image prompting spread through observation rather than documentation. For builders, this is a quiet thesis: agent literacy may be a social skill, not a technical one.

Meanwhile, the technical scaffolding keeps getting denser. Jason Liu, who works on developer experience at OpenAI, described his daily setup as a main agent with a /goal, running 3 sub-agents each with their own /goal. This recursive-goal pattern echoes what orchestration frameworks are formalizing — a recent DEV Community breakdown argues that simple chains hit a "stateless wall" in production, and that stateful orchestration is now table stakes for anything beyond demo-ware.

On the enterprise side, OpenHands launched an Agent Control Plane in early May 2026, pitched at organizations watching agents sprawl across repos and business units. KPMG's enterprise strategy guide, published the same week, lands in the same neighborhood — advising CIOs to apply a "structured readiness lens" before deploying, not after agents are already loose in the org.

For builders, the takeaway is that the next year of agent work is operational, not model-driven. The interesting questions are about goal decomposition, state management, and observability — not whether the underlying model can reason. For buyers, control planes look like an emerging procurement category, the way Kubernetes became one a decade ago. Worth watching: swyx asked publicly what percentage of Codex users have actually enabled the high-autonomy mode — a useful proxy for real agent adoption versus assistant-style use.

The bottom line: If you're building with agents in 2026, your bottleneck is probably orchestration, not intelligence — invest accordingly.

Sources