---
schema: "swft.publication/v1"
id: "agent-swarm-scale"
title: "Agent swarm scale: documented runs and adoption"
description: "Compare documented agent swarm scale with clear units, dates, outcomes, and limits: concurrent workers, cumulative runs, output, adoption, runtime, and tokens."
summary: "Agent scale has several meanings: workers active together, sessions accumulated over time, output produced, and systems used across a company. Compare documented runs using the same unit and time window, then inspect the result and the human effort behind it."
canonical: "https://swft.io/scale"
author: "SWFT Editorial"
author_type: "Organization"
published: "2026-09-08"
modified: "2026-09-08"
kind: "reference"
section: "References"
tags: ["agent swarms", "multi-agent systems", "software factory metrics", "AI adoption", "agent orchestration"]
evidence_labels: ["INFERENCE", "SELF-REPORT"]
source_ids: ["industry-map-bmw-review", "industry-map-bny-q1", "itau-devin-customer", "scale-anthropic-compiler", "scale-cognition-annual-review", "scale-cursor-parallel-agents", "scale-cursor-self-driving", "scale-kimi-agent-swarm", "scale-klarna-alphaevolve", "scale-kosmos-paper", "scale-openai-navier-stokes", "scale-openai-research-usage", "vercel-introducing-eve"]
authorship_disclosure: "AI-drafted from the cited public sources and independently checked by a second AI editorial-review agent (Codex) for source fit, claim boundaries, overlap, and reader utility. SWFT Editorial is responsible for corrections."
---

# Agent swarm scale: documented runs and adoption

What public numbers reveal about concurrent agents, repeated attempts, useful output, and adoption across organizations.

> **Authorship:** AI-drafted from the cited public sources and independently checked by a second AI editorial-review agent (Codex) for source fit, claim boundaries, overlap, and reader utility. SWFT Editorial is responsible for corrections.

## Quick answer

Agent scale has several meanings: workers active together, sessions accumulated over time, output produced, and systems used across a company. Compare documented runs using the same unit and time window, then inspect the result and the human effort behind it.

An **agent swarm** coordinates agents toward a shared task. A **fleet** can contain agents doing separate jobs. Some public research accounts report thousands of simultaneous workers; organizational adoption figures often count something else.

Browse the dated first-party reports below, checked by Codex (AI research agent) on September 8, 2026. Each preserves the unit, reporting window, outcome and limits. Research runs, deployed systems and product capacities are labeled separately. This selected record is not an industry-wide adoption survey.

## Agents working at once

Concurrency counts active workers at the same time. A reported capacity is a supported ceiling; a peak says how high one run reached. Neither establishes the average or the useful work produced.

| System | Reported scale | Measurement window | Primary source |
| --- | --- | --- | --- |
| Anthropic · C compiler experiment | 16 parallel agents | Two-week experiment reported February 2026 | [Anthropic](https://www.anthropic.com/engineering/building-c-compiler) |
| Cursor · Browser research harness | Hundreds of concurrent workers | Nearly one week, reported January 2026 | [Cursor](https://cursor.com/blog/scaling-agents) |
| Moonshot AI · Kimi K2.5 Agent Swarm | Up to 100 subagents | K2.5 research preview; blog publication date not displayed | [Moonshot AI](https://www.kimi.com/en/blog/kimi-k2-5) |
| OpenAI · Navier–Stokes research agents | Approximately 10,000 concurrent agents | September 1–5, 2026; 88 hours from initial launches | [OpenAI](https://openai.com/index/navier-stokes-solution/) |

### Anthropic: concurrent agents

**C compiler experiment: 16 parallel agents.** Claude instances working on one shared compiler project.

Produced a compiler that built Linux 6.9 for three architectures.

**Limits:** Human-designed tests and interventions shaped the work. GCC handled the x86 16-bit boot phase; the compiler remained incomplete.

Domain: Software development. Setting: research. Basis: reported run. Source published: 2026-02-05. Checked by Codex (AI research agent): 2026-09-08.

### Cursor: concurrent agents

**Browser research harness: Hundreds of concurrent workers.** Workers contributing to the same experimental browser codebase.

The team reported more than a million lines across roughly a thousand files.

**Limits:** No exact concurrent count in this account. Code volume does not establish browser completeness or production readiness.

Domain: Software development. Setting: research. Basis: reported run. Source published: 2026-01-14. Checked by Codex (AI research agent): 2026-09-08.

### Moonshot AI: subagent capacity

**Kimi K2.5 Agent Swarm: Up to 100 subagents.** Announced parallel swarm capacity, with up to 1,500 tool calls.

Moonshot reports execution up to 4.5 times faster than its single-agent baseline.

**Limits:** Capacity and benchmark claims. This source does not establish a customer run with 100 simultaneous agents or disclose typical utilization.

Domain: Research and office work. Setting: research. Basis: reported capacity. Source published: not publicly stated. Checked by Codex (AI research agent): 2026-09-08.

### OpenAI: concurrent agents

**Navier–Stokes research agents: Approximately 10,000 concurrent agents.** The group credited with the proposed proof; other groups explored related problems.

OpenAI reports 17 additional hours for formalization and verification in Lean using GPT-6 Astra.

**Limits:** Not independent mathematical acceptance. People redirected work and consolidated findings. Sustained concurrency and cost were not disclosed.

Domain: Mathematics. Setting: research. Basis: reported run. Source published: 2026-09-08. Checked by Codex (AI research agent): 2026-09-08.

## Sessions and search attempts

A session or rollout is one stretch of agent work. A candidate program is one possible solution. These counts accumulate across a project, including retries and rejected work; compare the units before comparing the numbers.

| System | Reported scale | Measurement window | Primary source |
| --- | --- | --- | --- |
| Anthropic · C compiler experiment | Nearly 2,000 sessions | Two weeks | [Anthropic](https://www.anthropic.com/engineering/building-c-compiler) |
| Edison Scientific / FutureHouse · Kosmos | Over 200 agent rollouts | Per research run, up to 12 hours | [Edison Scientific and collaborators](https://arxiv.org/abs/2511.02824) |
| Klarna · AlphaEvolve | Nearly 6,000 candidate programs | Three-week training optimization project | [Klarna Engineering](https://engineering.klarna.com/beyond-prompting-how-algorithmic-evolution-doubled-our-training-speed-8f874af3080d) |

### Anthropic: Claude Code sessions

**C compiler experiment: Nearly 2,000 sessions.** Successive sessions across the 16-worker project.

Reported consumption: 2 billion input tokens, 140 million output tokens, and just under $20,000 in API charges.

**Limits:** Sessions accumulated over time. API charges exclude the researcher's labor and do not establish total project cost.

Domain: Software development. Setting: research. Basis: reported run. Source published: 2026-02-05. Checked by Codex (AI research agent): 2026-09-08.

### Edison Scientific / FutureHouse: agent rollouts

**Kosmos: Over 200 agent rollouts.** Literature search and data analysis coordinated through a shared world model.

Average run: 1,500 papers read and 42,000 lines of analysis code executed. Scientist evaluation judged 79.4% of report statements accurate.

**Limits:** Rollouts are cumulative. Concurrent worker count is not reported. The six-month human-equivalent claim is collaborators' estimation, not a controlled comparison.

Domain: Scientific research. Setting: research. Basis: reported run. Source published: 2025-11-04. Checked by Codex (AI research agent): 2026-09-08.

### Klarna: candidate programs

**AlphaEvolve: Nearly 6,000 candidate programs.** Search over code for one model's training pipeline, with reproducibility constraints.

Training throughput rose from 49 to roughly 97 samples per second under those constraints.

**Limits:** Candidate count is not agent concurrency. A faster, nondeterministic candidate was rejected; one experiment continued through 631 evaluations without improvement. Deployment coverage and search cost were not quantified.

Domain: Financial services. Setting: pilot. Basis: reported run. Source published: 2026-03-30. Checked by Codex (AI research agent): 2026-09-08.

## Output over time

Throughput connects output to a time window. Commits, proposed changes, and merged changes represent different stages of delivery. A peak hourly rate cannot stand in for a sustained rate.

| System | Reported scale | Measurement window | Primary source |
| --- | --- | --- | --- |
| Cursor · Browser research harness | Approximately 1,000 commits/hour at peak | Peak within a one-week run | [Cursor](https://cursor.com/blog/self-driving-codebases) |

### Cursor: commits per hour

**Browser research harness: Approximately 1,000 commits/hour at peak.** Research harness also accumulated 10 million tool calls during the week.

Agents repeatedly contributed changes to an experimental browser.

**Limits:** The system tolerated temporary errors. Commit rate does not measure accepted features, defect-free releases, or sustained throughput.

Domain: Software development. Setting: research. Basis: reported run. Source published: 2026-02-05. Checked by Codex (AI research agent): 2026-09-08.

## Use across an organization

Deployed systems and accepted work show where agents have entered an organization. They do not reveal how many agents collaborate on any one task or what share of all suitable work is automated.

| System | Reported scale | Measurement window | Primary source |
| --- | --- | --- | --- |
| Cognition customers · Devin | Hundreds of thousands of merged PRs | First 18 months after launch, reported November 2025 | [Cognition](https://cognition.com/blog/devin-annual-performance-review-2025) |
| Vercel · eve | More than 100 production agents | Internal fleet described June 2026 | [Vercel](https://vercel.com/blog/introducing-eve) |
| BNY · Eliza digital employees | Approximately 140 digital employees | First-quarter 2026 results | [BNY](https://www.bny.com/assets/corporate/documents/pdf/investor-relations/earnings/quarterly-update-presentation-1q-2026.pdf) |
| BMW · CodeRabbit code review | More than 1,000 developers supported | August 2026; collaboration began more than two years earlier | [BMW Group](https://www.press.bmwgroup.com/usa/article/detail/T0460064EN_US/bmw-i-ventures-in-coderabbit-to-advance-independent-ai-review-in-software-development) |
| Itaú · Devin | 75% of teams use Devin | Undated customer account; measurement period not stated | [Cognition / Devin](https://devin.ai/customers/itau) |

### Cognition customers: merged pull requests

**Devin: Hundreds of thousands of merged PRs.** Vendor aggregate across thousands of customer companies.

Documents accepted code changes across the customer base.

**Limits:** Cognition also reports a 67% merge rate, without underlying attempt counts or a clear cohort window. Shared-task concurrency and production outcomes are not established.

Domain: Software development across industries. Setting: production. Basis: organization total. Source published: 2025-11-14. Checked by Codex (AI research agent): 2026-09-08.

### Vercel: deployed agents

**eve: More than 100 production agents.** Distinct agents serving roles across Vercel's business.

The source describes recurring use in analysis, support, sales, and content work.

**Limits:** Deployed agents can have separate tasks and schedules. The figure does not establish simultaneous activity or a hundred-agent swarm on one task.

Domain: Business operations. Setting: production. Basis: organization total. Source published: 2026-06-17. Checked by Codex (AI research agent): 2026-09-08.

### BNY: deployed multi-agent solutions

**Eliza digital employees: Approximately 140 digital employees.** BNY's term for multi-agent systems working alongside human colleagues.

A deployment snapshot within the bank's AI program, reported separately from roughly 220 enterprise AI solutions in production.

**Limits:** Each digital employee can contain multiple agents. This is neither a concurrent-worker count nor a measure of completed work per system.

Domain: Financial services. Setting: production. Basis: organization total. Source published: 2026-04-16. Checked by Codex (AI research agent): 2026-09-08.

### BMW: developers supported

**CodeRabbit code review: More than 1,000 developers supported.** BMW's deployed source-code review system, supporting vehicle software development.

Reviews proposed changes for developers deciding what is ready to ship.

**Limits:** Support coverage is not active-use frequency, agent count, or an end-to-end software factory. BMW-specific review volume and defect rates are not disclosed.

Domain: Automotive software. Setting: production. Basis: organization total. Source published: 2026-08-12. Checked by Codex (AI research agent): 2026-09-08.

### Itaú: share of teams

**Devin: 75% of teams use Devin.** Vendor-hosted account featuring Itaú technology leaders.

Documents organizational adoption; the account describes code maintenance and repair workflows.

**Limits:** The total number of teams, definition of use, and reporting window are missing. It does not mean 75% of work is automated or reveal concurrency.

Domain: Financial services. Setting: production. Basis: organization total. Source published: not publicly stated. Checked by Codex (AI research agent): 2026-09-08.

## Runtime and inference

Tokens measure model input or output; runtime measures elapsed agent activity. Track both with the task and outcome. Different models, tools, and accounting methods can produce very different totals for similar work.

| System | Reported scale | Measurement window | Primary source |
| --- | --- | --- | --- |
| OpenAI · Navier–Stokes research agents | Approximately 130 billion output tokens | September 2026 Navier–Stokes effort | [OpenAI](https://openai.com/index/navier-stokes-solution/) |
| OpenAI · Internal coding agents | 3.1 agent-workdays per human workday | Research organization, mid-August 2026 | [OpenAI](https://openai.com/index/research-acceleration-view-inside-openai/) |

### OpenAI: output tokens

**Navier–Stokes research agents: Approximately 130 billion output tokens.** The Navier–Stokes effort, which also produced 2.7 million messages. No per-group breakdown is given.

Cumulative activity associated with the proposed proof.

**Limits:** The wider multi-problem campaign used about 300 billion output tokens. Neither total measures discoveries or simultaneous agents; cost is undisclosed.

Domain: Mathematics. Setting: research. Basis: reported run. Source published: 2026-09-08. Checked by Codex (AI research agent): 2026-09-08.

### OpenAI: runtime ratio

**Internal coding agents: 3.1 agent-workdays per human workday.** Aggregate agent runtime normalized to an eight-hour workday.

Shows how much agent activity had entered daily research work.

**Limits:** Runtime is not equivalent productive labor. Coverage is incomplete. Separately, January–July task analysis found interventions in more than half of successful tasks estimated at 4–8 human hours.

Domain: AI research. Setting: production. Basis: organization total. Source published: 2026-09-06. Checked by Codex (AI research agent): 2026-09-08.

## How to compare a new claim

Start with the unit and the time window. Then ask what entered the system, what counted as accepted work, and how much human effort was needed. Keep failed attempts, repair, and review in the calculation. “Not publicly stated” is a useful answer when the source omits a field.

For a claim about concurrency, look for average utilization as well as a peak. For a claim about adoption, look for the denominator: eligible teams, suitable tasks, or total accepted changes. For a speed claim, keep the quality and reproducibility requirements fixed. A run that finds more candidates may still spend most of its time revisiting weak ideas.

This collection is a selected public record. Companies disclose different things, and some disclose very little. Absence here does not establish absence of use. Old observations remain useful when their measurement window stays visible; a newly checked source does not turn an old figure into a current one.

The [software factory metrics guide](/software-factory-metrics) explains how to measure useful output, review effort, and total cost within your own team. The [company cases](/companies) explain the workflows behind the numbers.

## How we know

- **First-party report (SELF-REPORT)** The dated observations reproduce metric definitions and limitations from the organizations reporting the work; they do not establish independent comparative performance. Sources: [Building a C compiler with a team of parallel Claudes](https://www.anthropic.com/engineering/building-c-compiler); [Scaling long-running autonomous coding](https://cursor.com/blog/scaling-agents); [Kimi K2.5: Visual Agentic Intelligence](https://www.kimi.com/en/blog/kimi-k2-5); [On the Navier–Stokes Millennium Prize Problem](https://openai.com/index/navier-stokes-solution/); [Kosmos: An AI Scientist for Autonomous Discovery](https://arxiv.org/abs/2511.02824); [Beyond Prompting: How Algorithmic Evolution Doubled our Training Speed](https://engineering.klarna.com/beyond-prompting-how-algorithmic-evolution-doubled-our-training-speed-8f874af3080d); [Towards self-driving codebases](https://cursor.com/blog/self-driving-codebases); [Devin's 2025 Performance Review: Learnings From 18 Months of Agents At Work](https://cognition.com/blog/devin-annual-performance-review-2025); [Introducing eve](https://vercel.com/blog/introducing-eve); [First Quarter 2026 Financial Results, page 4](https://www.bny.com/assets/corporate/documents/pdf/investor-relations/earnings/quarterly-update-presentation-1q-2026.pdf); [BMW i Ventures invests in CodeRabbit to Advance Independent AI Review in Software Development](https://www.press.bmwgroup.com/usa/article/detail/T0460064EN_US/bmw-i-ventures-in-coderabbit-to-advance-independent-ai-review-in-software-development); [How Itaú Deployed AI Across the SDLC at Global-Finance Scale](https://devin.ai/customers/itau); [Research acceleration: The view inside OpenAI](https://openai.com/index/research-acceleration-view-inside-openai/).
- **Analysis (INFERENCE)** Concurrency, accumulated work, adoption, and inference consumption answer different questions. SWFT groups them by measurement type and retains missing evidence instead of ranking unlike systems.

## Sources

- **First-party report (SELF-REPORT)** [BMW i Ventures invests in CodeRabbit to Advance Independent AI Review in Software Development](https://www.press.bmwgroup.com/usa/article/detail/T0460064EN_US/bmw-i-ventures-in-coderabbit-to-advance-independent-ai-review-in-software-development) — BMW Group; published 2026-08-12; accessed 2026-09-08. BMW describes deployed code review supporting over 1,000 developers. Vendor-wide review volume and future repair capabilities are separate claims.
- **First-party report (SELF-REPORT)** [First Quarter 2026 Financial Results, page 4](https://www.bny.com/assets/corporate/documents/pdf/investor-relations/earnings/quarterly-update-presentation-1q-2026.pdf) — BNY; published 2026-04-16; accessed 2026-09-08. Q1 2026 deployment snapshot: approximately 220 enterprise AI solutions and 140 digital employees. These are solution counts, not simultaneous workers.
- **First-party report (SELF-REPORT)** [How Itaú Deployed AI Across the SDLC at Global-Finance Scale](https://devin.ai/customers/itau) — Cognition / Devin; accessed 2026-09-08. Undated vendor-hosted account featuring named Itaú operators; measurement windows are incomplete.
- **First-party report (SELF-REPORT)** [Building a C compiler with a team of parallel Claudes](https://www.anthropic.com/engineering/building-c-compiler) — Anthropic; published 2026-02-05; accessed 2026-09-08. Compiler experiment with distinct counts for parallel agents and cumulative sessions, plus costs, tokens, human work, and remaining limitations.
- **First-party report (SELF-REPORT)** [Devin's 2025 Performance Review: Learnings From 18 Months of Agents At Work](https://cognition.com/blog/devin-annual-performance-review-2025) — Cognition; published 2025-11-14; accessed 2026-09-08. Vendor account of aggregate merged pull requests across customers, rather than a single coordinated run.
- **First-party report (SELF-REPORT)** [Scaling long-running autonomous coding](https://cursor.com/blog/scaling-agents) — Cursor; published 2026-01-14; accessed 2026-09-08. Engineering account of concurrent workers on experimental software projects, including a browser.
- **First-party report (SELF-REPORT)** [Towards self-driving codebases](https://cursor.com/blog/self-driving-codebases) — Cursor; published 2026-02-05; accessed 2026-09-08. Follow-up explaining throughput, accumulated tool use, temporary errors, and the evolving research harness.
- **First-party report (SELF-REPORT)** [Kimi K2.5: Visual Agentic Intelligence](https://www.kimi.com/en/blog/kimi-k2-5) — Moonshot AI; accessed 2026-09-08. Technical blog reporting swarm capacity and benchmark latency; the retrieved page does not display a publication date.
- **First-party report (SELF-REPORT)** [Beyond Prompting: How Algorithmic Evolution Doubled our Training Speed](https://engineering.klarna.com/beyond-prompting-how-algorithmic-evolution-doubled-our-training-speed-8f874af3080d) — Klarna Engineering; published 2026-03-30; accessed 2026-09-08. Operator account of candidate-program search, measured training speed, reproducibility constraints, and a prolonged plateau.
- **First-party report (SELF-REPORT)** [Kosmos: An AI Scientist for Autonomous Discovery](https://arxiv.org/abs/2511.02824) — Edison Scientific and collaborators; published 2025-11-04; updated 2025-11-05; accessed 2026-09-08. System authors' research report on cumulative agent rollouts, analysis volume, report accuracy, and collaborators' estimates of human effort.
- **First-party report (SELF-REPORT)** [On the Navier–Stokes Millennium Prize Problem](https://openai.com/index/navier-stokes-solution/) — OpenAI; published 2026-09-08; accessed 2026-09-08. Research announcement disclosing coordination scale, inference volume, elapsed time, and human steering; mathematical acceptance is outside SWFT's review.
- **First-party report (SELF-REPORT)** [Research acceleration: The view inside OpenAI](https://openai.com/index/research-acceleration-view-inside-openai/) — OpenAI; published 2026-09-06; accessed 2026-09-08. Organization-level measures of coding-agent runtime, inference valued at API prices, task outcomes, and human intervention.
- **First-party report (SELF-REPORT)** [Introducing eve](https://vercel.com/blog/introducing-eve) — Vercel; published 2026-06-17; accessed 2026-09-01. Vercel's account of the lowercase eve framework, its internal production use, filesystem convention, durable runs, sandboxes, connections, traces, evals, previews, and named agents.

## Read next

- [Software factory metrics that matter](/software-factory-metrics)
- [How Cursor combines cloud coding agents, Bugbot review, and security agents](/companies/cursor-agent-factory)
- [Vercel eve: the shared infrastructure behind more than 100 production agents](/companies/vercel-eve)