Skip to content

Agent Sprawl: Unmanaged Sub-Agent and Skill Proliferation

An agent and skill catalog grows faster than it is pruned, leaving unowned overlapping entries that degrade routing accuracy.

Agent sprawl is the structural failure mode that follows agent-native development without governance. Augment Code defines it as "deploying agents without ownership, accountability, or governance" (Augment Code); IBM calls it "the uncontrolled proliferation of AI agents across an organization" (IBM Think). The shape is the same at every scale: the catalog gains entries faster than anyone prunes them.

Sprawl is not over-agentification. Over-agentification is a per-task design error — choosing an agent when a deterministic workflow would do (Augment Code). Sprawl is a fleet-management failure over time. Each agent may have been the right choice when created, but the catalog has no owner, no audit, and no deprecation path.

Symptoms

  • Ambiguous routing. Two sub-agents have overlapping descriptions, so the orchestrator picks unpredictably.
  • Orphaned agents. The original author has moved on, and nobody knows what the agent does or whether it can be deleted. Only 18% of organizations keep a current, complete inventory of their AI agents (IBM IBV via IBM Think).
  • Silent duplication. Two teams build near-identical agents without knowing the other exists. Salesforce's 2026 Connectivity Benchmark reports that 50% of enterprise AI agents operate in silos (IBM Think).
  • No deprecation path. Agents are created but never retired — what IBM calls "a decommissioning failure" (IBM Think).
  • Capability overlap by tool allowlist. Two agents request the same tool surface for the same job, which splits the audit trail.

Why it works (as a failure mode)

Three compounding mechanisms degrade the catalog faster than authors can keep up. Routing degradation comes first. Anthropic's tool-authoring guidance warns that "when tools overlap in function or have a vague purpose, agents can get confused about which ones to use," and that "too many tools or overlapping tools can distract agents from pursuing efficient strategies" (Anthropic). Ownership drift compounds it. Augment Code notes that "scaling multi-agent systems increases coordination drift as more roles, prompts, and handoffs accumulate" (Augment Code). Coordination cost is third: each new agent multiplies the handoff surface. Gartner reports that only 13% of organizations have the right AI agent governance, and projects that the average Fortune 500 will run over 150,000 agents by 2028 (Gartner via IBM Think).

Mitigations

These apply to a single repository, not just enterprise scale:

  • Named ownership. Every agent and skill carries a current owner in frontmatter, so orphans surface in audit.
  • Discoverability index. A /list-skills / /list-commands-style command exposes the catalog to people and the orchestrator, so duplicates show up rather than stay buried.
  • Periodic catalog audit. Review quarterly for agents never selected, duplicated tool allowlists, and overlapping descriptions.
  • Explicit deprecation. Retired agents are archived or deleted — IBM names lifecycle management as a primary control (IBM Think).
  • Standardize creation. When building a new agent is easier through the standard path than outside it, governance becomes self-reinforcing (IBM Think).

When this backfires

Calling proliferation "sprawl" too early suppresses the divergence a team needs to find its taxonomy:

  • Early-stage discovery (first ~6 months). Teams need to learn which tasks benefit from a dedicated agent, a workflow, or a shared skill. The Fortune 500-scale governance toolkit is wildly out of proportion for a five-person team running eight sub-agents.
  • Solo catalogs. Anthropic's routing mechanism is real at any size, but ownership drift needs multiple authors and time. A one-person catalog runs without a registry until it scales.
  • Intentional fork divergence. Two agents that look like duplicates may be deliberate forks for separate codebases or risk tiers — the Copy-Paste Agent page covers the same trade-off.
  • Generate-then-prune sprints. Specialization is not always beneficial — generalist agents can outperform specialists when concurrent execution improves throughput (Predicting Multi-Agent Specialization). Sprawl framing during the generation phase removes the variants needed for selection.

Key Takeaways

  • Sprawl is unmanaged proliferation, not proliferation itself — the failure is no owner, no audit, no deprecation path.
  • Distinct from over-agentification: a per-task design error vs. a fleet-management failure over time.
  • Three compounding mechanisms degrade the catalog: routing degradation, ownership drift, coordination cost growth.
  • Mitigations applicable in a single repo: named owners, discoverability index, periodic audit, explicit deprecation.
  • The framing fires only after the catalog stabilizes; during early discovery, governance overhead suppresses the divergence the team needs.
Emerging — single recent source Last reviewed: 2026-06-16