What is subagent sprawl?
Subagent sprawl is the build-up of many overlapping subagents inside one agent system, typically one per team or feature, with no shared standards or single owner, which adds latency and cost and makes behaviour inconsistent.
The term comes from Anthropic's "Building agents: when to fold subagents into your main agent loop" (October 2026). The article describes large organisations where each team ships its own subagent, a platform team builds an orchestrator on top, and the company ends up with hundreds of overlapping subagents that have no enforced standards and no specific owner.
It names a companion problem, the Frankenprompt: every team adds its own rules to the shared system prompt until the instructions conflict and the agent has no consistent voice. Both are framed as Conway's law at work, with the agent's architecture turning into a copy of the org chart.
Symptoms
Context lost between agents
Details drop out as work passes from one subagent to the next, which the article compares to a game of telephone.
Overlapping subagents
Several subagents can do roughly the same job, and the orchestrator picks between them inconsistently.
Conflicting instructions
The shared prompt carries rules from different teams that pull in opposite directions.
Local evals only
Each team scores its own subtask, so nobody owns end-to-end quality and leadership sees averages that hide real failures.
Added latency
Every extra hop is another model call. One company with about a dozen specialist subagents found the delay hurt the user experience and made an on-brand tone hard to keep.
Why it happens
The article traces sprawl to how work is divided. Subagents get split by problem type or by owning team rather than by what context each step actually needs, and teams treat subagents as the only way to isolate context, even though progressive disclosure of skills and tools now handles much of that inside one loop.
Independent research points the same way. Google Research's study of 180 agent configurations (January 2026) found that independent multi-agent systems amplified errors 17.2 times, against 4.4 times for systems with central coordination.
How to fold subagents back in
Build an end-to-end eval first
Even about 10 golden examples, covering each production agent chain and each rule the agent must not break, with every task run more than once.
Baseline with subagents as tools
The main agent owns the conversation and a shared response schema, and checks each result.
Merge tightly coupled pairs
Specialists that chain on each other's output, touch the same data or are called together go first.
Fold one at a time
Turn the subagent's prompts into skills and its API calls into tools, compare evals with and without it, and keep it only if the folded version scores worse.
One owner
A single central team owns the system prompt and end-to-end quality, and vets new tools, skills and evals before they ship.
What the folding buys
One customer that compared its multi-agent harness with a single agent loop saw roughly 2 times faster runtime. The article also reports that tool choice tends to degrade past 30 to 50 tool definitions, so tools should be grouped by user intent, and that in one test a folded subagent broke a process in only one of two runs, which is why each eval task needs more than one run.
Subagents still earn a place when a single loop measurably degrades from context bloat, when tasks truly need to run in parallel with fresh context, or when one agent cannot pick the right tools or carries conflicting instructions.
Signs of subagent sprawl and the usual fix
| What it looks like | Usual fix | |
|---|---|---|
| Overlap | Two subagents answer the same kind of request differently | Merge them, or fold both into skills on one loop |
| Frankenprompt | System prompt rules contradict each other | One owning team and a reviewed prompt |
| Telephone game | Facts from the first turn are missing by the third agent | Fewer hops, or subagents as tools with a shared schema |
| Local metrics | Each team's score looks fine while users complain | An end-to-end eval set owned centrally |
| Latency | Simple questions take several model calls | Route simple requests to a single loop |
Common questions
What is a Frankenprompt?
A system prompt stitched together from many teams' rules, so that instructions conflict and the agent loses a consistent voice. Anthropic's October 2026 guidance names it alongside subagent sprawl.
Are subagents a bad idea?
No. They help for context isolation, parallel work, verification and keeping permissions separate. Sprawl is the problem of having many overlapping ones with no owner, not of having any at all.
How do teams fix subagent sprawl?
By building an end-to-end eval, then folding subagents into one main loop one at a time as skills and tools, keeping a subagent only where the eval shows the folded version does worse.
Related
Subagent
A helper agent with a clean context that hands back a short answer.
Single Agent vs Multi-Agent
One loop with good tools, compared with several agents splitting the work.
Progressive disclosure
Show the menu first, load the dish only when it is ordered.
Agent Skills
Folders of instructions, scripts and resources that an agent loads only when the task needs them.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.