Gemini 4 Carbon (High) took first place in Text Arena with 1528 points, surpassing Gemini 4 Argon (High)'s 1520 points. It ranks second in Agent Arena, behind only Claude Opus 5.5, and has entered a new Pareto frontier; in Web Dev Arena it ranks fourth with 1768 points, two points behind third-place Claude Sonnet 5.5. Arena data states its median cost per Agent Arena task is $1.80, 72% lower than top-ranked Claude Opus 5.5 (Max). Its Artificial Analysis V5 score is 59.
DeepLearning.AI has launched a new Context Engineering course, taught by Chroma CEO Jeff Huber and CMPND CEO Drew Breunig, framing context engineering as the main thread for keeping Agents reliable and token-efficient in complex multi-step tasks. The course builds a harness from scratch in Python and uses traces and evals to quantify each change, covering three capabilities—Agentic Search, Context Management, and Agentic Memory—across four modules, with labs mostly consisting of about 1-hour coding exercises. The course assumes intermediate Python, familiarity with LLM applications and tool calling, and basic retrieval.
epLearning.AI has launched a new Context Engineering course, taught by Chroma CEO Jeff Huber and CMPND CEO Drew Breunig, framing context engineering as the main thread for keeping Agents reliable and token-efficient in complex multi-step tasks. The course builds a harness from scratch in Python and uses traces and evals to quantify each change, covering three capabilities—Agentic Search, Context Management, and Agentic Memory—across four modules, with labs mostly consisting of about 1-hour codin
Arena.ai says Gemini 4 Carbon ranks 2nd in the real-task-based Agent Arena with a net improvement of 16.2%, and sits on the Pareto frontier for cost, time, and output tokens per task. Compared with Gemini 4 Argon, which ranks 7th with an 11% improvement, Carbon is 5 places higher and 5.2 percentage points better; the gains mainly come from a +13.2 percentage point increase in the positive-feedback-to-complaint ratio, +5.6 percentage points in confirmed success, and +4.8 percentage points in steerability. Arena says Carbon's availability will be announced by Google DeepMind.
Arena.ai says Gemini 4 Carbon ranks 2nd in the real-task-based Agent Arena with a net improvement of 16.2%, and sits on the Pareto frontier for cost, time, and output tokens per task. Compared with Gemini 4 Argon, which ranks 7th with an 11% improvement, Carbon is 5 places higher and 5.2 percentage points better; the gains mainly come from a +13.2 percentage point increase in the positive-feedback-to-complaint ratio, +5.6 percentage points in confirmed success, and +4.8 percentage points in stee
Arena's empirical results, based on nearly 9,000 agent sessions, show that OpenAI's GPT-6.1 Terra (Max) ranks 8th in Agent Arena, with a net improvement of +10.6% and a median cost of $0.62 per task. This cost is 43% lower than the $1.09 of 7th-place Gemini 4 Argon (High), 30% lower than the $0.89 of 6th-place GPT-6.1 Sol (Max), and 88% lower than the $5.16 of 5th-place GPT-6 Astra (Max). In the subcategories, it ranks 3rd in Bash Reliability (+13.2%), 7th in Steerability (+13.3%), and 9th in Praise vs Complaint (+16.8%).
Arena's empirical results, based on nearly 9,000 agent sessions, show that OpenAI's GPT-6.1 Terra (Max) ranks 8th in Agent Arena, with a net improvement of +10.6% and a median cost of $0.62 per task. This cost is 43% lower than the $1.09 of 7th-place Gemini 4 Argon (High), 30% lower than the $0.89 of 6th-place GPT-6.1 Sol (Max), and 88% lower than the $5.16 of 5th-place GPT-6 Astra (Max). In the subcategories, it ranks 3rd in Bash Reliability (+13.2%), 7th in Steerability (+13.3%), and 9th in Pr
Matt Palmer demonstrated opening Chrome and real Slack on the Linux remote desktop of a Cursor Cloud Agent, having the Agent generate a Slack UI motion video with code, and then iterating and revising after checking easing and motion patterns against the real interface. He broke the process down into an inner loop (the Agent does the work, checks against the real thing, and fixes) and an outer loop (a human provides references, preferences, and permissions and accepts the output), and proposed the Do→Codify→Generalize ladder, packaging one success into a Skill and then generalizing it. He emphasized that Slack was not decompiled; he merely gave the Agent a reference and asked it to extract visual patterns by comparison.
Matt Palmer demonstrated opening Chrome and real Slack on the Linux remote desktop of a Cursor Cloud Agent, having the Agent generate a Slack UI motion video with code, and then iterating and revising after checking easing and motion patterns against the real interface. He broke the process down into an inner loop (the Agent does the work, checks against the real thing, and fixes) and an outer loop (a human provides references, preferences, and permissions and accepts the output), and proposed t
Arena's statistics on Agent Arena show that OpenAI's GPT-6.1 Sol (Max) costs $1.91 per task, 85% less than Claude Opus 5.5 (Max) at $12.51 and 87% less than GPT-6 Astra (Max) at $14.93. Arena notes that the sources of these two cost advantages differ: relative to Astra, Sol's advantage comes mainly from a lower token unit price; relative to Opus, it comes mainly from consuming fewer tokens per task. Cost per task is calculated by multiplying the number of input, output, and cache-read tokens by their respective unit prices, and Arena pinpoints the causes of cost reductions by comparing prices and token usage.
Arena's statistics on Agent Arena show that OpenAI's GPT-6.1 Sol (Max) costs $1.91 per task, 85% less than Claude Opus 5.5 (Max) at $12.51 and 87% less than GPT-6 Astra (Max) at $14.93. Arena notes that the sources of these two cost advantages differ: relative to Astra, Sol's advantage comes mainly from a lower token unit price; relative to Opus, it comes mainly from consuming fewer tokens per task. Cost per task is calculated by multiplying the number of input, output, and cache-read tokens by
The Hermes Agent open-sourced by Nous Research externalizes learning into files, using MEMORY.md and USER.md to store memory, SKILL.md to store reusable skills, and a wiki to store domain knowledge, while the underlying model weights remain unchanged. The article breaks down its closed-loop learning mechanism, including post-task background review that automatically creates and improves skills. The built-in llm-wiki skill 2.1.0 maintains the knowledge base with a three-layer structure of raw/, wiki, and SCHEMA.md, along with disciplines such as page-creation thresholds, bidirectional links, source-text hashing, and contradiction registration, and comes with an inspection mechanism.
The Hermes Agent open-sourced by Nous Research externalizes learning into files, using MEMORY.md and USER.md to store memory, SKILL.md to store reusable skills, and a wiki to store domain knowledge, while the underlying model weights remain unchanged. The article breaks down its closed-loop learning mechanism, including post-task background review that automatically creates and improves skills. The built-in llm-wiki skill 2.1.0 maintains the knowledge base with a three-layer structure of raw/, w
The author ran their own QQ Mail inbox of 5,039 emails through the email Agent on mobile; of 69 emails from the past month, about 20 needing action were filtered out in 43 seconds. The article provides prompts for 5 common scenarios, plus a memory template of "user profile + 10 behavioral guidelines," usable by copying without writing any code. Core scenarios include finding hidden AI auto-charges and SaaS renewals in 17 seconds, and merging multiple banks' Chinese and English monthly statements and calculating foreign currencies separately in 30 seconds.
The author ran their own QQ Mail inbox of 5,039 emails through the email Agent on mobile; of 69 emails from the past month, about 20 needing action were filtered out in 43 seconds. The article provides prompts for 5 common scenarios, plus a memory template of "user profile + 10 behavioral guidelines," usable by copying without writing any code. Core scenarios include finding hidden AI auto-charges and SaaS renewals in 17 seconds, and merging multiple banks' Chinese and English monthly statements
Blake Anderson has open-sourced "ai-app-playbook," an operating playbook for coding agents to execute, covering product selection, building, launch, and first-month profitability. It has two layers: a human-facing playbook/00–13 chapters and an Agent-facing AGENTS.md (Phase 0–14, with each step marked [AGENT] or [HUMAN]); before starting, one must pass the Gate A scoring. The tech stack is fixed on Cloudflare Workers, D1, Better Auth, OpenRouter routing to DeepSeek, Superwall plus Stripe, etc., and explicitly lists a long set of non-choices such as Vercel, Supabase, and Next.js.
Blake Anderson has open-sourced "ai-app-playbook," an operating playbook for coding agents to execute, covering product selection, building, launch, and first-month profitability. It has two layers: a human-facing playbook/00–13 chapters and an Agent-facing AGENTS.md (Phase 0–14, with each step marked [AGENT] or [HUMAN]); before starting, one must pass the Gate A scoring. The tech stack is fixed on Cloudflare Workers, D1, Better Auth, OpenRouter routing to DeepSeek, Superwall plus Stripe, etc.,
The Claude Blog discusses when to fold sub-agents into the main loop when building Agents, arguing that on the Claude Platform, a single main loop plus on-demand loading of Skills/tools often yields better quality and lower cost and latency across multiple internal comparisons. The article notes that one sub-agent per team plus aggregate orchestration often replicates the organizational chart, bringing problems such as context loss, Frankenprompts, and localized evals; one customer, when comparing a multi-Agent harness with a single loop, found the latter about 2× faster. Sub-agents are recommended only for scenarios such as parallelism, context bloat, or conflicts between tools and instructions, and the article provides a migration path of first building end-to-end evals and folding in one at a time.
The Claude Blog discusses when to fold sub-agents into the main loop when building Agents, arguing that on the Claude Platform, a single main loop plus on-demand loading of Skills/tools often yields better quality and lower cost and latency across multiple internal comparisons. The article notes that one sub-agent per team plus aggregate orchestration often replicates the organizational chart, bringing problems such as context loss, Frankenprompts, and localized evals; one customer, when compari
Kimi K3.1 Pro (Max) ranks 11th in Arena's Agent Arena, achieving a +7.6% net improvement based on 4K+ real agent sessions, up 6 percentage points from Kimi K3 (Max)'s 25th place. Its median cost per task is $1.03, 13% lower than the previous version, ranking 8th, 9th, and 10th in the Work, Code, and Chat categories respectively. Arena stated that the model's open weights will be released soon, at which point it will become the top-ranked open-source model in Agent Arena, surpassing DeepSeek-v4.1-Flash.
Kimi K3.1 Pro (Max) ranks 11th in Arena's Agent Arena, achieving a +7.6% net improvement based on 4K+ real agent sessions, up 6 percentage points from Kimi K3 (Max)'s 25th place. Its median cost per task is $1.03, 13% lower than the previous version, ranking 8th, 9th, and 10th in the Work, Code, and Chat categories respectively. Arena stated that the model's open weights will be released soon, at which point it will become the top-ranked open-source model in Agent Arena, surpassing DeepSeek-v4.1
Shao Meng compiled a workflow from Tibo: use Agents to design 3–4 AI creators with fixed faces and daily routines, segmented by buyer; each day they find viral hits, rewrite them in the character's voice, render with the same face, and schedule 24 hours in advance, with humans only reviewing the queue and weekly topic updates. The workflow has 9 steps and comes with 12 prompts, starting from connecting to Revid MCP and writing the brain.md master brain, through creating faces, setting topics, breaking down viral post structures, running daily batches, operating independent accounts, and weekly reviews. The article emphasizes that account bios must state AI creator, fake reviews and health/medical claims are prohibited, products should occupy only one beat of the video, and content is only scheduled rather than posted directly to leave room for human review.
Shao Meng compiled a workflow from Tibo: use Agents to design 3–4 AI creators with fixed faces and daily routines, segmented by buyer; each day they find viral hits, rewrite them in the character's voice, render with the same face, and schedule 24 hours in advance, with humans only reviewing the queue and weekly topic updates. The workflow has 9 steps and comes with 12 prompts, starting from connecting to Revid MCP and writing the brain.md master brain, through creating faces, setting topics, br
Artificial Analysis released Coding Agent Index v2.0, replacing DeepSWE with Cognition's FrontierCode v1.1 and tightening repository Q&A scoring; the leaderboard evaluates model-plus-harness combinations. Currently first is Claude Code + Claude Opus 5.5 (max) at 61.7, followed by Claude Code + Sonnet 5.5 at 59.3 and Muse Code + Muse Wonder Full at 58.9; the first OpenAI combination, Codex + GPT-6 Astra, is at 54.7. For the same 21 configurations, the mean across the old and new leaderboards fell from 54.8 to 47.4, with 20 declining, mainly because FrontierCode's mean is about 45.1%, far below DeepSWE's 65.6%.
Artificial Analysis released Coding Agent Index v2.0, replacing DeepSWE with Cognition's FrontierCode v1.1 and tightening repository Q&A scoring; the leaderboard evaluates model-plus-harness combinations. Currently first is Claude Code + Claude Opus 5.5 (max) at 61.7, followed by Claude Code + Sonnet 5.5 at 59.3 and Muse Code + Muse Wonder Full at 58.9; the first OpenAI combination, Codex + GPT-6 Astra, is at 54.7. For the same 21 configurations, the mean across the old and new leaderboards fell
Aerial is a tool that can replace the default launcher on Chinese smart TVs without root. It uses a "launcher guard" to overlay the original launcher when it is woken up, and leverages Android notification listening to revive within 0.5 seconds after standby cleanup, so home-screen ads disappear, though boot ads still cannot be blocked. During deployment, connect the TV and computer to the same WiFi, enable ADB and find the TV's IP, then copy the prompt from the Aerial official website and hand it to Agents such as Doubao and WorkBuddy to execute; the user only needs to allow USB debugging in the TV pop-up, and the whole process takes about five minutes. The tool—from code and APK to the official website—was produced entirely through AI Vibe Coding, gained over 200 stars within one day of launching on GitHub, and subsequently added features such as icon hiding and adding signal input sources to the home page, along with adaptation for multiple TV models.
be blocked. During deployment, connect the TV and computer to the same WiFi, enable ADB and find the TV's IP, then copy the prompt from the Aerial official website and hand it to Agents such as Doubao and WorkBuddy to execute; the user only needs to allow USB debugging in the TV pop-up, and the whole process takes about five minutes. The tool—from code and APK to the official website—was produced entirely through AI Vibe Coding, gained over 200 stars within one day of launching on GitHub, and s
Browser Use has released its 2026 roundup of browser Agent MCP servers, dividing such servers into two modes—driver and delegate—and evaluating four of them one by one: Browser Use Cloud MCP, Browser Use open-source MCP, Hyperbrowser MCP, and agent-browser chat. In a comparative evaluation across 100 BU Bench V2.1 tasks, the author found that the four driver-type tools consumed a median of 2.1M to 6.7M input tokens per task, all of which count toward the coding Agent's context. The article also provides each server's installation commands, model selection, billing methods, and known limitations, and explains that long tasks, parallel tasks, and structured-output scenarios are better suited to the delegate mode.
Browser Use has released its 2026 roundup of browser Agent MCP servers, dividing such servers into two modes—driver and delegate—and evaluating four of them one by one: Browser Use Cloud MCP, Browser Use open-source MCP, Hyperbrowser MCP, and agent-browser chat. In a comparative evaluation across 100 BU Bench V2.1 tasks, the author found that the four driver-type tools consumed a median of 2.1M to 6.7M input tokens per task, all of which count toward the coding Agent's context. The article also
Lecture 9 of Stanford's CS 329Z course presents a selection diagram for Agent system optimization: the prerequisite for optimization is having data, a system, and evaluation metrics all in place; generally, context optimization (prompts and few-shots) is done first because it is cheap, fast, and sample-efficient; when context optimization is insufficient, or when you want to distill capabilities into a small model, or when prompts grow longer and longer, then you move on to changing the weights. The lecture notes also review reflective prompt optimization methods such as OPRO, MIPROv2, TextGrad, and GEPA; weight optimization paths such as SFT, RL (PPO vs. GRPO), and on-policy distillation; and approaches that let agents self-optimize, such as the Darwin Gödel Machine, Meta-Harness, and karpathy's autoresearch. They also point out that agents are already decent optimizers, making evaluation metrics the link that most needs to be controlled.
Lecture 9 of Stanford's CS 329Z course presents a selection diagram for Agent system optimization: the prerequisite for optimization is having data, a system, and evaluation metrics all in place; generally, context optimization (prompts and few-shots) is done first because it is cheap, fast, and sample-efficient; when context optimization is insufficient, or when you want to distill capabilities into a small model, or when prompts grow longer and longer, then you move on to changing the weights.
Cameron R. Wolfe, a senior research scientist at Netflix, has published a roadmap for evaluating subjective tasks, proposing a three-step process: human evaluation, offline automated evaluation, and online monitoring. The article presents statistical tools such as power analysis, confidence intervals, and repeated sampling, along with five criteria that evaluation sets must meet: diversity, realism, sufficient difficulty, discriminative power, and accuracy. The overall process is: offline evaluation screens ideas; promising versions are handed to human evaluation; after A/B testing, they are fully launched and continuously monitored, with the results feeding back into the evaluation set.
Cameron R. Wolfe, a senior research scientist at Netflix, has published a roadmap for evaluating subjective tasks, proposing a three-step process: human evaluation, offline automated evaluation, and online monitoring. The article presents statistical tools such as power analysis, confidence intervals, and repeated sampling, along with five criteria that evaluation sets must meet: diversity, realism, sufficient difficulty, discriminative power, and accuracy. The overall process is: offline evalua
