# AI Agent Swarms: Scaling Laws, Frameworks, and 10-Agent Caps

URL: https://3ammarketer.com/blog/ai-agent-swarms-scaling-laws-frameworks
Published: 2026-10-12

> AI agent swarms deliver diminishing returns past 10 agents, where 10x scaling returns 3x to 5x performance. Here is how production frameworks and benchmark economics compare.

Multi-agent AI swarms hit sharp diminishing returns beyond 10 workers, where scaling the agent pool by 10x yields only 3x to 5x higher performance instead of linear gains. While coordinating multiple agents improves pass rates on complex reasoning tasks, the bulk of execution capability stems from underlying foundation model power rather than decentralized orchestration overhead. [1][2][3]

## What an AI agent swarm actually is

An AI agent swarm distributes execution across specialized sub-agents that communicate, invoke tools, and transfer execution context dynamically. Rather than relying on a monolithic prompt to complete a composite problem in one continuous generation, swarms decompose instructions into targeted operational steps. [4][5]

In lightweight orchestration setups like OpenAI Swarm, execution relies on two primitives: agents and handoffs. Agents run stateless client-side calls directly via the Chat Completions API and do not utilize the Assistants API. In this architecture, an Agent dataclass defaults to model 'gpt-4o', instructions 'You are a helpful agent.', and an empty list of functions. [4]

Swarm converts agent instructions directly into the system prompt as the first message of a conversation. When an execution transfer occurs, an agent handoff swaps the system prompt while leaving the chat history intact. The core client.run() loop executes tool calls, appends results, switches the active agent if directed, and updates context variables until no function calls remain. [4]

## The scaling curve flattens past 10 agents

Research into multi-agent systems reveals that adding more parallel workers does not scale output proportionally. According to Toby Ord, scaling up agents in a swarm by 10x yields 10^lambda times the performance, which equals roughly 3x to 5x the output rather than the efficiency of allocating 10x as many tokens to an individual worker. [3]

This drop-off aligns with empirical findings from major frontier model evaluations. The Claude Opus 5.5 system card reported that the largest performance increase in multi-agent results occurred when scaling from 1 to 10 agents, with much smaller gains beyond that. In time-constrained scenarios, a chart from Anthropic showed 100 agents far outperforming 10 and 30 agents within a 2-hour window, though setups with 10 or 30 agents caught up when allowed more time. [1]

In a paper released September 17, researchers from Microsoft Research and UC Berkeley observed that collaborating agent swarms scored higher on certain benchmarks than individual agents with equal time and resources. Similarly, the paper 'More Agents Is All You Need' established an Agent Forest sampling-and-voting method showing LLM performance scales with the number of agents instantiated. [1][6]

In September, OpenAI announced that 10,000 agents collaborated to solve a famous math problem in just a few days. On September 8, OpenAI announced a proof for the Clay Institute's Navier-Stokes existence and smoothness problem after 88 hours of AI runtime, during which the swarm exchanged roughly 2.7 million messages. [1][2]

Examining those results, Noam Brown stated that he lacked evidence showing 10,000 agents were necessary to solve the Navier-Stokes problem, noting that 1,000 agents might have performed similarly. Brown attributed less than 10% of the credit for the breakthrough to the multi-agent approach, pointing instead to the underlying power of the trained model. Brown also observed on the Dwarkesh Podcast that previous reasoning models had been trained to think alone and that receiving messages from peer agents broke their concentration, even while describing agents talking during multi-agent training as the most 'feel the AGI' moment since reasoning models and chain of thought emerged. [1]

## Production frameworks in 2026: what replaces experimental repos

Teams deploying swarm architectures in production have largely moved away from educational repositories. OpenAI's Swarm repository, which requires Python 3.10 or higher, is officially replaced by the OpenAI Agents SDK for production workloads. Similarly, the Microsoft AutoGen repository has entered maintenance mode and is community-managed without receiving new features or enhancements, with the Microsoft Agent Framework designated as its enterprise-ready successor. [4][7]

![Notice on the Microsoft AutoGen repository announcing maintenance mode and migration to Microsoft Agent Framework.](https://dygolixznokmekyvajbb.supabase.co/storage/v1/object/public/images/content/3dc9a7939dfbca32707fc7aa.png)

![Notice on the OpenAI Swarm GitHub repository stating deprecation in favor of the OpenAI Agents SDK.](https://dygolixznokmekyvajbb.supabase.co/storage/v1/object/public/images/content/3acfb8ccdf2ff2a0f1f4acfe.png)

For complex task management, frameworks like CrewAI organize agents through structured workflows. CrewAI crews can execute processes using either sequential or hierarchical flows, where a designated manager coordinates execution. To prevent API saturation across multi-agent loops, CrewAI provides a max_rpm parameter that defines the maximum requests per minute across the entire execution, overriding individual agents' max_rpm settings. Practitioners comparing structural orchestration options can evaluate these mechanics alongside our guide to [building AI agents frameworks](/blog/building-ai-agents-frameworks). [8]

Real-world implementations also demonstrate how multi-agent structures scale dynamically from basic prompts. Prompting Codex with GPT-6 Astra Ultra to brainstorm post ideas automatically spun up 3 agents, and providing a 3-team framework expanded execution to 13 agents. Understanding these mechanics is essential when budgeting for [ai agents for business pricing breakdown](/blog/ai-agents-for-business-pricing-breakdown). [2]

## Benchmark task economics and token costs

Deploying multiple agents multiplies token consumption across every coordination cycle. On the SciUniverse benchmark of 92 tasks across 17 task families, Claude Fable 5.1 (xhigh) achieved a 45.3% pass rate at $40.61 per task. By comparison, GPT-5 Astra (xhigh) reached a 32.5% pass rate at $52.37 per task, while Claude Opus 5 (xhigh) achieved 30.5% at $46.31 per task. [3]

SciUniverse benchmark task performance and cost across 92 tasks. [3]

| Model Split | SciUniverse Pass Rate | Cost per Task |
| --- | --- | --- |
| Claude Fable 5.1 (xhigh) | 45.3% | $40.61 |
| GPT-5 Astra (xhigh) | 32.5% | $52.37 |
| Claude Opus 5 (xhigh) | 30.5% | $46.31 |

Cost per task on the SciUniverse benchmark in dollars. [3]

- Claude Fable 5.1 (xhigh): 40.61 $
- GPT-5 Astra (xhigh): 52.37 $
- Claude Opus 5 (xhigh): 46.31 $

SciUniverse benchmark pass rate across 92 tasks by model split. [3]

- Claude Fable 5.1 (xhigh): 45.3 %
- GPT-5 Astra (xhigh): 32.5 %
- Claude Opus 5 (xhigh): 30.5 %

For workloads demanding high parallelism, Kimi K2.5 by Moonshot AI can analyze a task, break it into pieces, and dynamically spawn up to 100 specialized sub-agents to execute sub-tasks simultaneously. Moonshot AI reports this parallelization cuts execution time by up to 4.5x for complex workflows. Kimi K2.5 was trained across 15 trillion tokens combining text and visual data, and its API pricing runs about 60 cents per million input tokens and roughly $2.50 per million output tokens. [5]

Token prices vary widely when orchestrating large agent clusters across different model classes. From our data, OpenAI API rates captured on 2026-10-03 show gpt-6-astra at $10 input and $50 output per million tokens, gpt-6.1-sol at $2 input and $10 output, and gpt-6-luna at $0.1 input and $0.5 output. For specialized models, gpt-5.6-sol costs $4 input and $20 output, while gpt-5.6-cyber costs $12.5 input and $75 output per million tokens. [ai marketing agents benchmark costs](/blog/ai-marketing-agents-benchmark-costs). [9]

From our data, for a workload processing 1 million input plus 1 million output tokens, OpenAI gpt-6-luna costs $0.6 while OpenAI gpt-5.6-cyber costs $87.5, which is 145.8 times as much. When swarms exchange millions of peer-to-peer messages, selecting higher-tier models without bounding agent count quickly inflates infrastructure spend. [9]

Limit: Published research does not prove that swarms beyond 10 agents improve final accuracy for tasks outside bounded reasoning or parallel search. Peer-to-peer message traffic frequently degrades reasoning concentration, and vendors do not publish standardized multi-agent coordination costs across identical tool interfaces.

## Deciding when to deploy multi-agent swarms

When building agentic workflows, limit swarm size to between 1 and 10 agents. Because a 10x expansion in agent headcount only yields 3x to 5x performance improvements, architectures scaling beyond 10 workers should only be used when execution latency is the primary constraint and sub-tasks can run purely in parallel. For sequential workflows, the sources point to the base model, not the agent count, as what moves results. [1][3]

## Sources

1. [Why agent swarms could be the next “scaling law”](https://www.understandingai.org/p/why-agent-swarms-could-be-the-next)
2. [The Dot and the Swarm - by Ethan Mollick - One Useful Thing](https://www.oneusefulthing.org/p/the-dot-and-the-swarm)
3. [Import AI 475: Swarm scaling; Google DeepMind watermarks biology; and the AI science economy](https://importai.substack.com/p/import-ai-475-swarm-scaling-google)
4. [GitHub - openai/swarm: Educational framework exploring ergonomic, lightweight multi-agent orchestration. Managed by OpenAI Solution team. · GitHub](https://github.com/openai/swarm)
5. [Agent Swarms: What It Means When AI Can Build Its Own Team](https://sidecar.ai/blog/agent-swarms-what-it-means-when-ai-can-build-its-own-team)
6. [[2402.05120] More Agents Is All You Need](https://arxiv.org/abs/2402.05120)
7. [GitHub - microsoft/autogen: A programming framework for agentic AI · GitHub](https://github.com/microsoft/autogen)
8. [Crews - CrewAI](https://docs.crewai.com/concepts/crews)
9. [LLM API pricing timeline (3AM Marketer data set)](https://3ammarketer.com/data/model-pricing)