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 as "the uncontrolled proliferation of AI agents across an organization" (IBM Think). The shape is the same at every scale: the catalog accumulates 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 suffices (Augment Code). Sprawl is a fleet-management failure over time: each individual agent may have been the right choice at creation, but the catalog has no owner, no audit, no deprecation path.

Symptoms

  • Ambiguous routing. Two sub-agents have overlapping descriptions and the orchestrator picks unpredictably.
  • Orphaned agents. The original author has moved on; nobody knows what the agent does or whether it can be deleted. Only 18% of organizations maintain a current and 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 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, splitting 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 "too many tools or overlapping tools can distract agents from pursuing efficient strategies" (Anthropic). Ownership drift compounds it — Augment Code notes "scaling multi-agent systems increases coordination drift as more roles, prompts, and handoffs accumulate" (Augment Code). Coordination cost is third: each new agent multiplies handoff surface. Gartner reports only 13% of organizations have the right AI agent governance, while projecting the average Fortune 500 will run over 150,000 agents by 2028 (Gartner via IBM Think).

Mitigations

Applicable to a single repository, not just enterprise scale:

  • Named ownership. Every agent and skill carries a current owner in frontmatter; orphans surface in audit.
  • Discoverability index. A /list-skills / /list-commands-style command exposes the catalog to humans and the orchestrator, so duplicates are visible rather than buried.
  • Periodic catalog audit. Quarterly review 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 discover which tasks benefit from a dedicated agent versus a workflow versus a shared skill. The Fortune 500-scale governance toolkit is wildly disproportionate to a five-person team running eight sub-agents.
  • Solo catalogs. Anthropic's routing mechanism is real at any size, but ownership drift requires 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 treats the same trade-off.
  • Generate-then-prune sprints. Specialization is not unconditionally 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
Feedback