Informe de harness
Quién corre el bucle, y cuánto le cuesta
El harness es el bucle alrededor del modelo: las herramientas, la política de contexto, los subagentes y el sandbox. Es donde la adopción se ve primero y donde se decide la cuenta. Esta página sigue cada semana las instalaciones de los principales harness y SDKs de agentes, y los estudios que miden cuánto le suma un harness al costo de una tarea.
Edición 2 · datos al 2026-10-16 · se actualiza cada semana · /api/harness-report
Esta edición
2026-10-16
Four new series from a research sweep on adoption, all with dated primary sources where they exist. Surveys: across three JetBrains waves the share of professional developers using Claude Code at work went from 3% to 18% to 39% in fourteen months while GitHub Copilot fell from 29% to 21% and Cursor returned to 12%; Stack Overflow's 2026 survey has Claude Code used by 65.5% of respondents in the past year against Copilot's 58.7%. Routers: OpenRouter's trailing-30-day top apps are Hermes Agent, Claude Code, Kilo Code and Cline. Companies: Cursor's stated run rate went from $1B in November 2025 to about $4B by June 2026, Lovable's from $100M to $600M in fourteen months, and Copilot's user count from 15M to 50M in fifteen months. Ten markers added.
- surveys.jetbrains: work adoption by tool across the 2025 State of Developer Ecosystem (Apr-Jun 2025), the January 2026 AI Pulse (10,000+ developers) and the 2026 State of Developer Ecosystem (May-Jul 2026, 15,000+): Claude Code 3 → 18 → 39%, Copilot 29 → 29 → 21%, Cursor 12 → 18 → 12%, Codex 0 → 3 → 16%, Antigravity 6 → 6%, OpenCode 7% (https://blog.jetbrains.com/research/2026/04/which-ai-coding-tools-do-developers-actually-use-at-work/; https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/). surveys.stackOverflow2026: coding agents used in the past year, n = 12,255, eighteen tools from Claude Code 65.5% to Aider 2.1% (https://survey.stackoverflow.co/2026/ai/data).
- topApps: OpenRouter's Top Apps by tokens, trailing 30 days through 15 Oct 2026: Hermes Agent 53.8T, Claude Code 34.3T, Kilo Code 24.3T, Cline 20.5T, pi 12.9T, Codex 11.7T, omp 10.6T, Freebuff 9.7T, OpenClaw 7.5T, DeepSeek Harness 6.1T, Command Code 4.2T, OpenHands 3.4T, with a 24 Aug 2026 snapshot alongside for comparison. Caveat: the coding apps roughly tripled and Codex and OpenClaw rose about a hundredfold between the two snapshots, which looks like an attribution or coverage change on OpenRouter's side rather than organic growth; treat the August column as a different measurement until OpenRouter explains (https://openrouter.ai/rankings; https://codesota.com/agentic/openrouter-apps).
- runRates: Cursor $100M (Jan 2025) → $500M (Jun 2025) → $1B (Nov 2025, cursor.com) → $2B (Feb 2026) → $3B (Apr 2026) → about $4B (Jun 2026), the last three via Bloomberg and TechCrunch sources; Lovable $100M (Jul 2025) → $600M (Sep 2026, lovable.dev); Claude Code over $500M (Sep 2025, press) → over $2.5B (Feb 2026, anthropic.com); Cognition $492M (May 2026, press) → almost $900M (Sep 2026, cognition.com). Press-sourced points carry secondary = true and hollow dots on the chart (sources in runRates.sources).
- users: GitHub Copilot 15M (Apr 2025) → 20M (Jul 2025) → 26M (Oct 2025) → 50M (Jul 2026) from Microsoft earnings calls; Antigravity weekly actives 1.5M (Feb 2026) → 2.4M (Jul 2026) from Alphabet calls; OpenAI Codex weekly users 2M (Mar 2026) → 4M (Apr) → 5M (Jun) → 7M (Jul), all via press because openai.com returned 403, and not extended past July because OpenAI then switched to a combined Codex plus ChatGPT Work count (15M in August, scope unstated) (sources in users.sources).
- Markers added: Copilot paid subscribers 4.7M (Jan 2026); one in three GitHub pull requests involves an agent (Jul 2026); Claude Code 4% of public GitHub commits (Feb 2026); JetBrains 90% weekly and 68% daily agent use; Stack Overflow 26.2% using agents at work with 63.1% daily; Requesty cost per Claude Code user $8 → $78 per month; OpenRouter: 6 Feb 2026 was the last day humans used more tokens than agents; Terminal-Bench 4.0 has 5 distinct harnesses across 35 rows; VS Code Marketplace installs; Lovable 60M+ projects.
- Not added, for lack of a primary or a definition: Codex counts after July 2026 (combined metric), Claude Code weekly-active counts (Anthropic says they doubled but gives no number), Kilo Code and OpenCode self-reported user totals (3M+ and 16M+ monthly, company claims with no basis stated), Devin user counts (aggregator only), and amplifying.ai agent-PR estimates (declared-metadata heuristics).
Agentes de código: descargas semanales
Descargas semanales de los harness de agentes de código instalables desde npm (Claude Code, Codex CLI, OpenCode, Gemini CLI) y PyPI (Aider, OpenHands), escala logarítmica. El historial de PyPI empieza en abril de 2026. Gemini CLI se retiró para los planes individuales el 18 de junio de 2026.
Descargas por semana
SDKs y frameworks de agentes: descargas semanales
Descargas semanales en PyPI de los frameworks sobre los que la gente construye sus propios harness, escala logarítmica. Los paquetes de npm de los mismos SDKs están en la tabla.
Descargas por semana
El impuesto del harness: mismo modelo, otra cuenta
Costo por intento en USD para el mismo modelo y benchmark en tres harness, escala logarítmica. El éxito difiere entre dos y cinco puntos; el costo difiere hasta cinco veces. Del estudio HarnessTax, septiembre de 2026.
Lo que cada harness manda antes de que digas nada
Tokens de entrada en la primera llamada al modelo para los mismos dos mensajes. El prompt de sistema y las definiciones de herramientas son el harness, no la tarea.
Tasa de acierto de caché por harness
Porción de tokens de entrada servidos desde la caché de prompt para el tráfico del enrutador Requesty, abril de 2026, según el harness que lo envió. La brecha entre el primero y el último es una diferencia de 2 a 4 veces en lo que cuesta el mismo trabajo.
Lo que los desarrolladores dicen que usan
Izquierda: porcentaje de desarrolladores profesionales que usan cada herramienta en el trabajo en tres encuestas de JetBrains (abr-jun 2025, ene 2026, may-jul 2026). Derecha: porcentaje de quienes respondieron la encuesta de Stack Overflow 2026 que usaron cada agente de código en el último año. Preguntas distintas, misma dirección.
JetBrains: usado en el trabajo
Stack Overflow 2026: usado en el último año
A dónde van los tokens del enrutador
Las apps con más tokens en OpenRouter en los últimos 30 días. Solo cuenta el tráfico que pasa por OpenRouter, así que los harness que llaman directo a los proveedores quedan subcontados; la tabla muestra la foto de finales de agosto para comparar.
Tokens, billones · trailing 30 days · 2026-10-15
Ingresos anualizados declarados
Ingresos anualizados declarados por la empresa (puntos llenos) o reportados por prensa con nombre (puntos huecos), escala logarítmica. Claude Code es una línea de producto dentro de Anthropic; las demás son empresas.
Ingresos anualizados, USD · los puntos huecos vienen de prensa
Usuarios reportados
Conteos de usuarios declarados en llamadas de resultados (puntos llenos) o relevados por prensa (puntos huecos), escala logarítmica. Las definiciones cambian por producto y están en la tabla; Codex se detiene en julio de 2026, cuando OpenAI pasó a un conteo combinado.
Usuarios, millones · los puntos huecos vienen de prensa
Marcadores de adopción
Cifras con fecha de publicaciones de empresas, comentarios de resultados y estudios publicados.
Antigravity weekly active users
2.4M
Jul 2026 · blog.google
Google ADK downloads, cumulative
~70M
Jul 2026 · blog.google
Gemini CLI at retirement for individual tiers: GitHub stars / merged PRs
100,000+ / 6,000
May 2026 · developers.googleblog.com
Opus share of Claude Code sessions vs chat
54% vs 10%
Jun 2026 · anthropic.com
OpenAI research org: agent-workdays per human workday
3.1
Aug 2026 · epoch.ai
Agentic vs human tokens on OpenRouter
~7x
Sep 2026 · openrouter.ai
GitHub Copilot paid subscribers
4.7M, up 75% YoY
Jan 2026 · microsoft.com
Pull requests on GitHub that involve an agent
1 in 3
Jul 2026 · microsoft.com
Claude Code share of public GitHub commits (SemiAnalysis, via Anthropic)
4%, doubled in a month
Feb 2026 · anthropic.com
Developers using coding agents at work at least weekly (JetBrains, 15,000+)
90%; 68% daily
Jul 2026 · blog.jetbrains.com
Respondents using AI agents at work; of those, daily (Stack Overflow 2026)
26.2%; 63.1% daily
Aug 2026 · survey.stackoverflow.co
Claude Code cost per user per month on Requesty
$8 (May 2025) → $78 (Apr 2026)
Apr 2026 · requesty.ai
Last day humans used more tokens than agents on OpenRouter
6 Feb 2026
Sep 2026 · openrouter.ai
Distinct harnesses on the Terminal-Bench 4.0 leaderboard
5 (35 rows)
Oct 2026 · tbench.ai
VS Code Marketplace installs: Copilot Chat / Claude Code / Codex / Cline
78.8M / 28.0M / 15.6M / 5.6M
16 Oct 2026 · marketplace.visualstudio.com
Lovable projects created
60M+ total; ~1M new per week
Aug 2026 · lovable.dev
Registro de cambios
Edición 2 · 2026-10-16
Four new series from a research sweep on adoption, all with dated primary sources where they exist. Surveys: across three JetBrains waves the share of professional developers using Claude Code at work went from 3% to 18% to 39% in fourteen months while GitHub Copilot fell from 29% to 21% and Cursor returned to 12%; Stack Overflow's 2026 survey has Claude Code used by 65.5% of respondents in the past year against Copilot's 58.7%. Routers: OpenRouter's trailing-30-day top apps are Hermes Agent, Claude Code, Kilo Code and Cline. Companies: Cursor's stated run rate went from $1B in November 2025 to about $4B by June 2026, Lovable's from $100M to $600M in fourteen months, and Copilot's user count from 15M to 50M in fifteen months. Ten markers added.
- surveys.jetbrains: work adoption by tool across the 2025 State of Developer Ecosystem (Apr-Jun 2025), the January 2026 AI Pulse (10,000+ developers) and the 2026 State of Developer Ecosystem (May-Jul 2026, 15,000+): Claude Code 3 → 18 → 39%, Copilot 29 → 29 → 21%, Cursor 12 → 18 → 12%, Codex 0 → 3 → 16%, Antigravity 6 → 6%, OpenCode 7% (https://blog.jetbrains.com/research/2026/04/which-ai-coding-tools-do-developers-actually-use-at-work/; https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/). surveys.stackOverflow2026: coding agents used in the past year, n = 12,255, eighteen tools from Claude Code 65.5% to Aider 2.1% (https://survey.stackoverflow.co/2026/ai/data).
- topApps: OpenRouter's Top Apps by tokens, trailing 30 days through 15 Oct 2026: Hermes Agent 53.8T, Claude Code 34.3T, Kilo Code 24.3T, Cline 20.5T, pi 12.9T, Codex 11.7T, omp 10.6T, Freebuff 9.7T, OpenClaw 7.5T, DeepSeek Harness 6.1T, Command Code 4.2T, OpenHands 3.4T, with a 24 Aug 2026 snapshot alongside for comparison. Caveat: the coding apps roughly tripled and Codex and OpenClaw rose about a hundredfold between the two snapshots, which looks like an attribution or coverage change on OpenRouter's side rather than organic growth; treat the August column as a different measurement until OpenRouter explains (https://openrouter.ai/rankings; https://codesota.com/agentic/openrouter-apps).
- runRates: Cursor $100M (Jan 2025) → $500M (Jun 2025) → $1B (Nov 2025, cursor.com) → $2B (Feb 2026) → $3B (Apr 2026) → about $4B (Jun 2026), the last three via Bloomberg and TechCrunch sources; Lovable $100M (Jul 2025) → $600M (Sep 2026, lovable.dev); Claude Code over $500M (Sep 2025, press) → over $2.5B (Feb 2026, anthropic.com); Cognition $492M (May 2026, press) → almost $900M (Sep 2026, cognition.com). Press-sourced points carry secondary = true and hollow dots on the chart (sources in runRates.sources).
- users: GitHub Copilot 15M (Apr 2025) → 20M (Jul 2025) → 26M (Oct 2025) → 50M (Jul 2026) from Microsoft earnings calls; Antigravity weekly actives 1.5M (Feb 2026) → 2.4M (Jul 2026) from Alphabet calls; OpenAI Codex weekly users 2M (Mar 2026) → 4M (Apr) → 5M (Jun) → 7M (Jul), all via press because openai.com returned 403, and not extended past July because OpenAI then switched to a combined Codex plus ChatGPT Work count (15M in August, scope unstated) (sources in users.sources).
- Markers added: Copilot paid subscribers 4.7M (Jan 2026); one in three GitHub pull requests involves an agent (Jul 2026); Claude Code 4% of public GitHub commits (Feb 2026); JetBrains 90% weekly and 68% daily agent use; Stack Overflow 26.2% using agents at work with 63.1% daily; Requesty cost per Claude Code user $8 → $78 per month; OpenRouter: 6 Feb 2026 was the last day humans used more tokens than agents; Terminal-Bench 4.0 has 5 distinct harnesses across 35 rows; VS Code Marketplace installs; Lovable 60M+ projects.
- Not added, for lack of a primary or a definition: Codex counts after July 2026 (combined metric), Claude Code weekly-active counts (Anthropic says they doubled but gives no number), Kilo Code and OpenCode self-reported user totals (3M+ and 16M+ monthly, company claims with no basis stated), Devin user counts (aggregator only), and amplifying.ai agent-PR estimates (declared-metadata heuristics).
Edición 1 · 2026-10-16
First edition. Weekly download series for nineteen harness and agent-framework packages from the npm and PyPI registries (a year of npm history, 26 weeks of PyPI), the HarnessTax cost-per-attempt table, the Portkey first-call context comparison, the Requesty cache-hit table by harness, and six dated adoption markers. Downloads are the only series here that updates itself; everything else moves when a study or a company publishes.
- Downloads: Codex CLI 23.9M and Claude Code 12.5M weekly npm downloads in the week of 5 Oct 2026; Claude Agent SDK 12.1M on npm and 7.6M on PyPI; LangGraph 12.2M on PyPI; AWS Strands 8.8M; OpenCode 3.0M; Google ADK 2.2M; OpenHands 190K; Aider 58K; Gemini CLI 407K after its June retirement for individual tiers (https://api.npmjs.org/downloads/range/last-year/@openai/codex; https://pypistats.org/api/packages/langgraph/overall).
- HarnessTax (Sep 2026): GPT-5.6 Luna $0.030 per attempt in Pi vs $0.152 in Claude Code on SWE-bench Lite at 53.3% vs 55.6% success; Fable 5 $0.666 (Pi) vs $0.890 (Codex) vs $1.329 (Claude Code) at 96.7% to 97.8%; GPT-5.6 Sol on Terminal-Bench 2.0 $0.421 in Pi at 83.3% vs $0.976 in Codex at 78.9%. The site renders only a loader; figures confirmed from the study's own write-up as reproduced by digitalapplied.com (https://arena.ai/blog/coding-agents-harness-tax/).
- First-call context for the same two messages: Pi about 2,600 tokens, Codex CLI about 15,000, Claude Code about 27,000 (Portkey, 13 Apr 2026, https://portkey.ai/blog/the-agent-harness-tax). Cache-hit rate by harness on Requesty traffic, Apr 2026: Claude Code 91.9%, OpenCode 89.0%, Aider 84.0%, Zed 80.1%, Roo Code 73.6%, Forge 63.9%, Cline 61.4%, Kilo Code 45.5% (https://requesty.ai).
Método: las descargas vienen de las APIs públicas de npm y PyPI, agrupadas en semanas ISO que empiezan el lunes, sin las semanas parciales; una descarga es una instalación o una actualización, no un usuario, y los conteos de npm incluyen tráfico de CI, así que lee la tendencia y no el nivel. Las cifras de costo y contexto se citan de los estudios nombrados en cada gráfica. Los marcadores son declaraciones con fecha de publicaciones o reportes de empresas. Se edita a mano; un registro de cambios anota cada modificación.