Businesses scaling their artificial intelligence deployments face a stark architectural reckoning as operating expenses mount. According to Anthropic, companies should evaluate whether they can consolidate sub-agents into a single agent that loads context on demand. The company's experience shows that many specialized sub-agents lead to higher latency, rising costs, and information loss at every handoff.
Organizations routinely replicate their corporate hierarchy in software, building multi-level flotillas of narrow sub-agents that eventually mirror their organizational charts. As Anthropic describes, this structural duplication generates Conway's Law for agents, where every internal team ships its own sub-agent until hundreds of overlapping modules pile up without shared standards.
The resulting operational friction manifests as bloated system prompts and severe performance degradation. What follows is what Anthropic calls a Frankenprompt, where every team adds rules to the system prompt that end up contradicting each other. A study by Vals found that agent teams can be faster than single agents but do not deliver higher quality, while scaling from ten to 30 or 100 agents brought no quality gains and significantly higher costs.
As OpenAI developer Eric Provencher has also observed regarding agent swarms, such setups are almost certainly wasted money. Beneath the surface of these multi-agent designs lies a direct financial drain on enterprise budgets, as redundant processing loops consume compute resources without yielding proportional output value.
Monolithic Efficiency and On-Demand Context
To curb these escalating infrastructure costs, market participants are pivoting toward centralized architectures powered by advanced foundation models. Newer models starting with Opus 4.5, along with Opus 5.5 and Fable 5.1, can manage complex plans directly in a single loop. In one comparison, Anthropic notes that the single-agent architecture ran about twice as fast as the multi-agent setup.
By keeping execution within a single loop and utilizing dynamic tool retrieval, applications eliminate the intermediate communication layers that previously drove up compute wait times. As Anthropic highlights, the company cut the Claude Code system prompt for newer models by over 80 percent, partly because skills and tools now load only when needed.
Anthropic cut the Claude Code system prompt for newer models by over 80 percent, partly because skills and tools now load only when needed.
This reduction in prompt overhead directly translates into lower inference costs and faster response times for enterprise end users. Instead of fetching context through code and injecting it directly into prompts, systems now rely on knowledge bases that the agent queries only when needed through a search tool.
Anthropic notes that the model's tool selection tends to degrade beyond 30 to 50 tool definitions, making strict governance of skill libraries mandatory. A small central team must own the system prompt and overall quality, while designated owners on each team maintain the skill libraries to keep them from ballooning.
Migration Path and Pragmatic Guardrails
Transitioning away from distributed agent swarms requires systematic architectural auditing rather than reckless consolidation. Anthropic has published a migration kit on GitHub to help teams make the transition. However, the vendor warns against merging blindly, since in one test with a support agent, folding in a sub-agent did reduce latency but broke a previously working process because the primary agent started answering on its own instead of consulting the right specialist.
Sub-agents retain a valid engineering purpose only under specific operational constraints. As Anthropic clarifies, sub-agents still make sense when a single agent's context window shows measurable performance degradation, when multiple tasks need to run in parallel with separate context windows, or when a single agent struggles with tool selection.
Editorial analysis
Anthropic is urging enterprises to consolidate specialized sub-agents into a single agent that loads context dynamically as needed. The company found that multi-agent pipelines trigger latency spikes, drive up compute costs, and drop critical context during handoffs. Back in late April we reported on Anthropic tracking enterprise ROI and labor shifts using a December dataset of 81,000 detailed responses on hiring and workplace roles. How much enterprise budget was burned on overengineered multi-agent workflows before the vendor issued this pivot?
We covered this: March 27, April 26