{"schema":"swft.agent-scale/v1","canonical":"https://swft.io/scale","modified":"2026-09-08","disclosure":"AI-drafted by SWFT Editorial. Measurements are dated first-party reports, not independent comparisons or an industry adoption census.","license":"https://creativecommons.org/licenses/by/4.0/","observations":[{"id":"anthropic-compiler-parallel-workers","organization":"Anthropic","system":"C compiler experiment","domain":"Software development","setting":"research","basis":"reported-run","measureKind":"concurrent-agents","displayValue":"16 parallel agents","unit":"concurrent agents","measurementWindow":"Two-week experiment reported February 2026","scope":"Claude instances working on one shared compiler project.","outcome":"Produced a compiler that built Linux 6.9 for three architectures.","limitations":"Human-designed tests and interventions shaped the work. GCC handled the x86 16-bit boot phase; the compiler remained incomplete.","sourceId":"scale-anthropic-compiler","sourceDate":"2026-02-05","checkedOn":"2026-09-08"},{"id":"cursor-browser-concurrent-workers","organization":"Cursor","system":"Browser research harness","domain":"Software development","setting":"research","basis":"reported-run","measureKind":"concurrent-agents","displayValue":"Hundreds of concurrent workers","unit":"concurrent agents","measurementWindow":"Nearly one week, reported January 2026","scope":"Workers contributing to the same experimental browser codebase.","outcome":"The team reported more than a million lines across roughly a thousand files.","limitations":"No exact concurrent count in this account. Code volume does not establish browser completeness or production readiness.","sourceId":"scale-cursor-parallel-agents","sourceDate":"2026-01-14","checkedOn":"2026-09-08"},{"id":"moonshot-swarm-capacity","organization":"Moonshot AI","system":"Kimi K2.5 Agent Swarm","domain":"Research and office work","setting":"research","basis":"reported-capacity","measureKind":"concurrent-agents","displayValue":"Up to 100 subagents","unit":"subagent capacity","measurementWindow":"K2.5 research preview; blog publication date not displayed","scope":"Announced parallel swarm capacity, with up to 1,500 tool calls.","outcome":"Moonshot reports execution up to 4.5 times faster than its single-agent baseline.","limitations":"Capacity and benchmark claims. This source does not establish a customer run with 100 simultaneous agents or disclose typical utilization.","sourceId":"scale-kimi-agent-swarm","sourceDate":null,"checkedOn":"2026-09-08"},{"id":"openai-navier-stokes-concurrency","organization":"OpenAI","system":"Navier–Stokes research agents","domain":"Mathematics","setting":"research","basis":"reported-run","measureKind":"concurrent-agents","displayValue":"Approximately 10,000 concurrent agents","unit":"concurrent agents","measurementWindow":"September 1–5, 2026; 88 hours from initial launches","scope":"The group credited with the proposed proof; other groups explored related problems.","outcome":"OpenAI reports 17 additional hours for formalization and verification in Lean using GPT-6 Astra.","limitations":"Not independent mathematical acceptance. People redirected work and consolidated findings. Sustained concurrency and cost were not disclosed.","sourceId":"scale-openai-navier-stokes","sourceDate":"2026-09-08","checkedOn":"2026-09-08"},{"id":"anthropic-compiler-sessions","organization":"Anthropic","system":"C compiler experiment","domain":"Software development","setting":"research","basis":"reported-run","measureKind":"cumulative-runs","displayValue":"Nearly 2,000 sessions","unit":"Claude Code sessions","measurementWindow":"Two weeks","scope":"Successive sessions across the 16-worker project.","outcome":"Reported consumption: 2 billion input tokens, 140 million output tokens, and just under $20,000 in API charges.","limitations":"Sessions accumulated over time. API charges exclude the researcher's labor and do not establish total project cost.","sourceId":"scale-anthropic-compiler","sourceDate":"2026-02-05","checkedOn":"2026-09-08"},{"id":"kosmos-research-rollouts","organization":"Edison Scientific / FutureHouse","system":"Kosmos","domain":"Scientific research","setting":"research","basis":"reported-run","measureKind":"cumulative-runs","displayValue":"Over 200 agent rollouts","unit":"agent rollouts","measurementWindow":"Per research run, up to 12 hours","scope":"Literature search and data analysis coordinated through a shared world model.","outcome":"Average run: 1,500 papers read and 42,000 lines of analysis code executed. Scientist evaluation judged 79.4% of report statements accurate.","limitations":"Rollouts are cumulative. Concurrent worker count is not reported. The six-month human-equivalent claim is collaborators' estimation, not a controlled comparison.","sourceId":"scale-kosmos-paper","sourceDate":"2025-11-04","checkedOn":"2026-09-08"},{"id":"klarna-training-candidates","organization":"Klarna","system":"AlphaEvolve","domain":"Financial services","setting":"pilot","basis":"reported-run","measureKind":"cumulative-runs","displayValue":"Nearly 6,000 candidate programs","unit":"candidate programs","measurementWindow":"Three-week training optimization project","scope":"Search over code for one model's training pipeline, with reproducibility constraints.","outcome":"Training throughput rose from 49 to roughly 97 samples per second under those constraints.","limitations":"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.","sourceId":"scale-klarna-alphaevolve","sourceDate":"2026-03-30","checkedOn":"2026-09-08"},{"id":"cursor-browser-commit-rate","organization":"Cursor","system":"Browser research harness","domain":"Software development","setting":"research","basis":"reported-run","measureKind":"throughput","displayValue":"Approximately 1,000 commits/hour at peak","unit":"commits per hour","measurementWindow":"Peak within a one-week run","scope":"Research harness also accumulated 10 million tool calls during the week.","outcome":"Agents repeatedly contributed changes to an experimental browser.","limitations":"The system tolerated temporary errors. Commit rate does not measure accepted features, defect-free releases, or sustained throughput.","sourceId":"scale-cursor-self-driving","sourceDate":"2026-02-05","checkedOn":"2026-09-08"},{"id":"cognition-merged-pr-adoption","organization":"Cognition customers","system":"Devin","domain":"Software development across industries","setting":"production","basis":"organization-total","measureKind":"adoption","displayValue":"Hundreds of thousands of merged PRs","unit":"merged pull requests","measurementWindow":"First 18 months after launch, reported November 2025","scope":"Vendor aggregate across thousands of customer companies.","outcome":"Documents accepted code changes across the customer base.","limitations":"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.","sourceId":"scale-cognition-annual-review","sourceDate":"2025-11-14","checkedOn":"2026-09-08"},{"id":"vercel-eve-deployed-agents","organization":"Vercel","system":"eve","domain":"Business operations","setting":"production","basis":"organization-total","measureKind":"adoption","displayValue":"More than 100 production agents","unit":"deployed agents","measurementWindow":"Internal fleet described June 2026","scope":"Distinct agents serving roles across Vercel's business.","outcome":"The source describes recurring use in analysis, support, sales, and content work.","limitations":"Deployed agents can have separate tasks and schedules. The figure does not establish simultaneous activity or a hundred-agent swarm on one task.","sourceId":"vercel-introducing-eve","sourceDate":"2026-06-17","checkedOn":"2026-09-08"},{"id":"bny-digital-employees","organization":"BNY","system":"Eliza digital employees","domain":"Financial services","setting":"production","basis":"organization-total","measureKind":"adoption","displayValue":"Approximately 140 digital employees","unit":"deployed multi-agent solutions","measurementWindow":"First-quarter 2026 results","scope":"BNY's term for multi-agent systems working alongside human colleagues.","outcome":"A deployment snapshot within the bank's AI program, reported separately from roughly 220 enterprise AI solutions in production.","limitations":"Each digital employee can contain multiple agents. This is neither a concurrent-worker count nor a measure of completed work per system.","sourceId":"industry-map-bny-q1","sourceDate":"2026-04-16","checkedOn":"2026-09-08"},{"id":"bmw-review-developer-reach","organization":"BMW","system":"CodeRabbit code review","domain":"Automotive software","setting":"production","basis":"organization-total","measureKind":"adoption","displayValue":"More than 1,000 developers supported","unit":"developers supported","measurementWindow":"August 2026; collaboration began more than two years earlier","scope":"BMW's deployed source-code review system, supporting vehicle software development.","outcome":"Reviews proposed changes for developers deciding what is ready to ship.","limitations":"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.","sourceId":"industry-map-bmw-review","sourceDate":"2026-08-12","checkedOn":"2026-09-08"},{"id":"itau-devin-team-adoption","organization":"Itaú","system":"Devin","domain":"Financial services","setting":"production","basis":"organization-total","measureKind":"adoption","displayValue":"75% of teams use Devin","unit":"share of teams","measurementWindow":"Undated customer account; measurement period not stated","scope":"Vendor-hosted account featuring Itaú technology leaders.","outcome":"Documents organizational adoption; the account describes code maintenance and repair workflows.","limitations":"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.","sourceId":"itau-devin-customer","sourceDate":null,"checkedOn":"2026-09-08"},{"id":"openai-navier-stokes-output-tokens","organization":"OpenAI","system":"Navier–Stokes research agents","domain":"Mathematics","setting":"research","basis":"reported-run","measureKind":"workload","displayValue":"Approximately 130 billion output tokens","unit":"output tokens","measurementWindow":"September 2026 Navier–Stokes effort","scope":"The Navier–Stokes effort, which also produced 2.7 million messages. No per-group breakdown is given.","outcome":"Cumulative activity associated with the proposed proof.","limitations":"The wider multi-problem campaign used about 300 billion output tokens. Neither total measures discoveries or simultaneous agents; cost is undisclosed.","sourceId":"scale-openai-navier-stokes","sourceDate":"2026-09-08","checkedOn":"2026-09-08"},{"id":"openai-research-agent-runtime","organization":"OpenAI","system":"Internal coding agents","domain":"AI research","setting":"production","basis":"organization-total","measureKind":"workload","displayValue":"3.1 agent-workdays per human workday","unit":"runtime ratio","measurementWindow":"Research organization, mid-August 2026","scope":"Aggregate agent runtime normalized to an eight-hour workday.","outcome":"Shows how much agent activity had entered daily research work.","limitations":"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.","sourceId":"scale-openai-research-usage","sourceDate":"2026-09-06","checkedOn":"2026-09-08"}],"sources":[{"id":"scale-anthropic-compiler","title":"Building a C compiler with a team of parallel Claudes","publisher":"Anthropic","url":"https://www.anthropic.com/engineering/building-c-compiler","published":"2026-02-05","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"Compiler experiment with distinct counts for parallel agents and cumulative sessions, plus costs, tokens, human work, and remaining limitations."},{"id":"scale-cursor-parallel-agents","title":"Scaling long-running autonomous coding","publisher":"Cursor","url":"https://cursor.com/blog/scaling-agents","published":"2026-01-14","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"Engineering account of concurrent workers on experimental software projects, including a browser."},{"id":"scale-kimi-agent-swarm","title":"Kimi K2.5: Visual Agentic Intelligence","publisher":"Moonshot AI","url":"https://www.kimi.com/en/blog/kimi-k2-5","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"Technical blog reporting swarm capacity and benchmark latency; the retrieved page does not display a publication date."},{"id":"scale-openai-navier-stokes","title":"On the Navier–Stokes Millennium Prize Problem","publisher":"OpenAI","url":"https://openai.com/index/navier-stokes-solution/","published":"2026-09-08","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"Research announcement disclosing coordination scale, inference volume, elapsed time, and human steering; mathematical acceptance is outside SWFT's review."},{"id":"scale-kosmos-paper","title":"Kosmos: An AI Scientist for Autonomous Discovery","publisher":"Edison Scientific and collaborators","url":"https://arxiv.org/abs/2511.02824","published":"2025-11-04","updated":"2025-11-05","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"System authors' research report on cumulative agent rollouts, analysis volume, report accuracy, and collaborators' estimates of human effort."},{"id":"scale-klarna-alphaevolve","title":"Beyond Prompting: How Algorithmic Evolution Doubled our Training Speed","publisher":"Klarna Engineering","url":"https://engineering.klarna.com/beyond-prompting-how-algorithmic-evolution-doubled-our-training-speed-8f874af3080d","published":"2026-03-30","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"Operator account of candidate-program search, measured training speed, reproducibility constraints, and a prolonged plateau."},{"id":"scale-cursor-self-driving","title":"Towards self-driving codebases","publisher":"Cursor","url":"https://cursor.com/blog/self-driving-codebases","published":"2026-02-05","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"Follow-up explaining throughput, accumulated tool use, temporary errors, and the evolving research harness."},{"id":"scale-cognition-annual-review","title":"Devin's 2025 Performance Review: Learnings From 18 Months of Agents At Work","publisher":"Cognition","url":"https://cognition.com/blog/devin-annual-performance-review-2025","published":"2025-11-14","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"Vendor account of aggregate merged pull requests across customers, rather than a single coordinated run."},{"id":"vercel-introducing-eve","title":"Introducing eve","publisher":"Vercel","url":"https://vercel.com/blog/introducing-eve","published":"2026-06-17","accessed":"2026-09-01","evidenceLabel":"SELF-REPORT","primary":true,"note":"Vercel's account of the lowercase eve framework, its internal production use, filesystem convention, durable runs, sandboxes, connections, traces, evals, previews, and named agents."},{"id":"industry-map-bny-q1","title":"First Quarter 2026 Financial Results, page 4","publisher":"BNY","url":"https://www.bny.com/assets/corporate/documents/pdf/investor-relations/earnings/quarterly-update-presentation-1q-2026.pdf","published":"2026-04-16","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"Q1 2026 deployment snapshot: approximately 220 enterprise AI solutions and 140 digital employees. These are solution counts, not simultaneous workers."},{"id":"industry-map-bmw-review","title":"BMW i Ventures invests in CodeRabbit to Advance Independent AI Review in Software Development","publisher":"BMW Group","url":"https://www.press.bmwgroup.com/usa/article/detail/T0460064EN_US/bmw-i-ventures-in-coderabbit-to-advance-independent-ai-review-in-software-development","published":"2026-08-12","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"BMW describes deployed code review supporting over 1,000 developers. Vendor-wide review volume and future repair capabilities are separate claims."},{"id":"itau-devin-customer","title":"How Itaú Deployed AI Across the SDLC at Global-Finance Scale","publisher":"Cognition / Devin","url":"https://devin.ai/customers/itau","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"Undated vendor-hosted account featuring named Itaú operators; measurement windows are incomplete."},{"id":"scale-openai-research-usage","title":"Research acceleration: The view inside OpenAI","publisher":"OpenAI","url":"https://openai.com/index/research-acceleration-view-inside-openai/","published":"2026-09-06","accessed":"2026-09-08","evidenceLabel":"SELF-REPORT","primary":true,"note":"Organization-level measures of coding-agent runtime, inference valued at API prices, task outcomes, and human intervention."}]}