Search “Agents”

Arena.ai publishes per-task cost comparison of models on Agent Arena

Arena.ai published a per-task cost comparison of models on Agent Arena on October 30, 2026. Statistics show that OpenAI's GPT-6.1 Sol (Max) costs $1.91 per task, 85% lower than Claude Opus 5.5 (Max) at $12.51 and 87% lower than GPT-6 Astra (Max) at $14.93. Arena.ai noted that these two cost advantages come from different sources: relative to Astra, Sol's advantage mainly comes from a lower per-token price; relative to Opus, it mainly comes from consuming fewer tokens per task. Per-task cost is calculated by multiplying the number of input, output, and cache-read tokens by their respective unit prices, and Arena.ai identifies the reasons for the cost reduction by comparing prices and token usage.

View story:Arena.ai Publishes Per-Task Cost Comparison of Models on Agent Arena

AYi's real-world test used an email Agent to organize 5,039 emails in QQ Mail

On October 27, 2026, AYi published a real-world test record, stating that on mobile it used an email Agent to handle its own 5,039 emails in QQ Mail: of the 69 emails from the past month, about 20 that needed handling were filtered out, taking 43 seconds. The post also provides instructions for 5 common scenarios and a memory template of "user profile + 10 behavioral guidelines," stating that they can be copied and used without writing any code; the core scenarios include finding hidden AI automatic deductions and SaaS renewals in 17 seconds, and merging Chinese and English monthly statements from multiple banks and separately calculating foreign currencies in 30 seconds. Currently, this real-world test appears only in one source, AYi AI Notes, and there have been no subsequent developments.

Larsen Cundric proposes defining the verification method before coding with an Agent

Larsen Cundric proposed a rule for coding with agents: before requesting a change, first agree on how both sides will confirm that the change works. The acceptance methods he gives include a test that currently fails, a command you can run yourself, and logs or screenshots of real run results; a "done" without evidence will be sent back. He believes most failed agent sessions begin with vague requirements and missing checks, and he invites readers to share their own criteria for trusting an agent's "done."

LangChain open-sources design Agent Content Studio

LangChain has open-sourced its design agent Content Studio and the harness that supports it. According to the LangChain Blog, the agent was built by LangChain's GTM engineering team in collaboration with the design team on Managed Deep Agents, runs inside Slack, and can generate videos, social graphics, UI mockups, and slide decks; it has already produced more than 3,000 assets and over 80 slide decks. The article breaks down its four-layer knowledge structure (skills, wiki, code validation, memory), the per-Slack-thread LangSmith Sandbox, the two rendering engines HyperFrames and the in-house motion-studio, and how the main agent orchestrates asynchronous sub-agents. The team says that after migrating to Managed Deep Agents, its own code shrank by about 80%, and the median model cost per request fell from about $8 in mid-September to about 50 cents. After open-sourcing, users can fork it and attach their own branding, then deploy it to their team's Slack via Managed Deep Agents.

LangChain open-sources its design Agent

LangChain has open-sourced its design agent Content Studio and the harness that supports it. According to the LangChain Blog, the agent was built by LangChain's GTM engineering team in collaboration with the design team on Managed Deep Agents, runs inside Slack, and can generate videos, social graphics, UI mockups, and slide decks; it has already produced more than 3,000 assets and over 80 slide decks. The article breaks down its four-layer knowledge structure (skills, wiki, code validation, memory), the per-Slack-thread LangSmith Sandbox, the two rendering engines HyperFrames and the in-house motion-studio, and how the main agent orchestrates asynchronous sub-agents. The team says that after migrating to Managed Deep Agents, its own code shrank by about 80%, and the median model cost per request fell from about $8 in mid-September to about 50 cents. After open-sourcing, users can fork it and attach their own branding, then deploy it to their team's Slack via Managed Deep Agents.

View story:LangChain open-sources design Agent Content Studio

AYi shares a real-world test of organizing 5,039 emails with an email Agent

On October 27, 2026, AYi published a real-world test record, stating that on mobile it used an email Agent to handle its own 5,039 emails in QQ Mail: of the 69 emails from the past month, about 20 that needed handling were filtered out, taking 43 seconds. The post also provides instructions for 5 common scenarios and a memory template of "user profile + 10 behavioral guidelines," stating that they can be copied and used without writing any code; the core scenarios include finding hidden AI automatic deductions and SaaS renewals in 17 seconds, and merging Chinese and English monthly statements from multiple banks and separately calculating foreign currencies in 30 seconds. Currently, this real-world test appears only in one source, AYi AI Notes, and there have been no subsequent developments.

View story:AYi's real-world test used an email Agent to organize 5,039 emails in QQ Mail

Larsen Cundric proposes defining verification methods before coding with an Agent

Larsen Cundric proposed a rule for coding with agents: before requesting a change, first agree on how both sides will confirm that the change works. The acceptance methods he gives include a test that currently fails, a command you can run yourself, and logs or screenshots of real run results; a "done" without evidence will be sent back. He believes most failed agent sessions begin with vague requirements and missing checks, and he invites readers to share their own criteria for trusting an agent's "done."

View story:Larsen Cundric proposes defining the verification method before coding with an Agent

LangChain summarizes the core components of a company-level Agent harness

On October 25, 2026, Harrison Chase and Caspar Broekhuizen of LangChain published a post summarizing the core components of a company-level Agent harness. Using LangChain's three internal agents for marketing, sales, and engineering—GTM Agent, OpenSWE, and Content Studio—as examples, the article distills five shared components: agents are placed in existing workplaces such as Slack; a distinction is made between acting on behalf of users and a fixed identity; global and user-level memory are set up as two tiers; code is executed within a sandbox; and many runs are triggered by schedules or events. The article also explains how the three handle identity, memory, and code execution specifically, and notes that all three are built on Managed Deep Agents. A subsequent Chinese-language report recounted the five common foundations described above and added that Managed Deep Agents is layered on top of the open-source Deep Agents harness, preserving room to customize agent behavior. At present, this summary still takes the official LangChain blog post as its original source.

LangChain summarizes the core components of an enterprise-grade Agent harness

On October 25, 2026, Harrison Chase and Caspar Broekhuizen of LangChain published a post summarizing the core components of a company-level Agent harness. Using LangChain's three internal agents for marketing, sales, and engineering—GTM Agent, OpenSWE, and Content Studio—as examples, the article distills five shared components: agents are placed in existing workplaces such as Slack; a distinction is made between acting on behalf of users and a fixed identity; global and user-level memory are set up as two tiers; code is executed within a sandbox; and many runs are triggered by schedules or events. The article also explains how the three handle identity, memory, and code execution specifically, and notes that all three are built on Managed Deep Agents. A subsequent Chinese-language report recounted the five common foundations described above and added that Managed Deep Agents is layered on top of the open-source Deep Agents harness, preserving room to customize agent behavior. At present, this summary still takes the official LangChain blog post as its original source.

View story:LangChain summarizes the core components of a company-level Agent harness

Shao Meng shares Claude blog: when to fold sub-agents into the main loop

On October 25, 2026, Shao Meng shared a Claude blog post's view on when to fold sub-agents into the main loop when building Agents. The blog argues that on the Claude Platform, a single main loop plus on-demand loading of Skills/tools often yields better quality, lower cost, and lower latency across multiple internal comparisons; one sub-agent per team plus aggregate orchestration often replicates the organizational chart, bringing problems such as context loss, Frankenprompts, and localized evals, and when one customer compared a multi-agent harness with a single loop, the latter was about 2× faster. The blog suggests reserving sub-agents only for scenarios such as parallelism, context bloat, or conflicts between tools and instructions, and offers a migration path of first building end-to-end evals and folding one at a time.

Shao Meng shares Claude blog: when to fold sub-agents into the main loop

On October 25, 2026, Shao Meng shared a Claude blog post's view on when to fold sub-agents into the main loop when building Agents. The blog argues that on the Claude Platform, a single main loop plus on-demand loading of Skills/tools often yields better quality, lower cost, and lower latency across multiple internal comparisons; one sub-agent per team plus aggregate orchestration often replicates the organizational chart, bringing problems such as context loss, Frankenprompts, and localized evals, and when one customer compared a multi-agent harness with a single loop, the latter was about 2× faster. The blog suggests reserving sub-agents only for scenarios such as parallelism, context bloat, or conflicts between tools and instructions, and offers a migration path of first building end-to-end evals and folding one at a time.

View story:Shao Meng shares Claude blog: when to fold sub-agents into the main loop

Shao Meng introduces Tibo's nine-step process for building an AI UGC creator team with Agents

On October 24, 2026, Shao Meng compiled and introduced a process from Tibo: using Agents to build your own AI UGC creator matrix. The process is designed by buyer segment, with 3–4 AI creators who have fixed faces and daily routines, completing each day the tasks of finding viral hits, rewriting them in the character's voice, rendering with the same face, and scheduling 24 hours in advance; humans only review the queue and weekly topic updates. The process has 9 steps in total and comes with 12 prompts, starting from connecting to the Revid MCP and writing the brain.md master brain, through creating faces, setting topics, breaking down viral structures, running daily batches, operating independent accounts, and weekly reviews. The article emphasizes that account bios should state AI creator, that fake reviews and health/medical claims are prohibited, that the product should occupy only one beat of the video, and that content is only scheduled rather than published directly so as to leave room for human review. There are currently no subsequent developments.

Arena ranks Kimi K3.1 Pro (Max) No. 11 in Agent Arena

On October 24, 2026, Arena added Kimi K3.1 Pro (Max) to the Agent Arena rankings, placing it at No. 11. The ranking is based on 4K multi-real-agent sessions, achieving a +7.6% net improvement, up 6 percentage points from Kimi K3 (Max)'s No. 25. Its median cost per task is $1.03, 13% lower than the previous version; it ranks No. 8, No. 9, and No. 10 in the Work, Code, and Chat categories, respectively. Arena said 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. There are currently no subsequent developments.

Arena ranks Kimi K3.1 Pro (Max) at No. 11 in Agent Arena

On October 24, 2026, Arena added Kimi K3.1 Pro (Max) to the Agent Arena rankings, placing it at No. 11. The ranking is based on 4K multi-real-agent sessions, achieving a +7.6% net improvement, up 6 percentage points from Kimi K3 (Max)'s No. 25. Its median cost per task is $1.03, 13% lower than the previous version; it ranks No. 8, No. 9, and No. 10 in the Work, Code, and Chat categories, respectively. Arena said 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. There are currently no subsequent developments.

View story:Arena ranks Kimi K3.1 Pro (Max) No. 11 in Agent Arena

Shao Meng introduces Tibo's nine-step process for building an AI UGC creator team with Agents

On October 24, 2026, Shao Meng compiled and introduced a process from Tibo: using Agents to build your own AI UGC creator matrix. The process is designed by buyer segment, with 3–4 AI creators who have fixed faces and daily routines, completing each day the tasks of finding viral hits, rewriting them in the character's voice, rendering with the same face, and scheduling 24 hours in advance; humans only review the queue and weekly topic updates. The process has 9 steps in total and comes with 12 prompts, starting from connecting to the Revid MCP and writing the brain.md master brain, through creating faces, setting topics, breaking down viral structures, running daily batches, operating independent accounts, and weekly reviews. The article emphasizes that account bios should state AI creator, that fake reviews and health/medical claims are prohibited, that the product should occupy only one beat of the video, and that content is only scheduled rather than published directly so as to leave room for human review. There are currently no subsequent developments.

View story:Shao Meng introduces Tibo's nine-step process for building an AI UGC creator team with Agents

Artificial Analysis publishes AA-AgentPerf results for Muse Spark Open

Artificial Analysis on October 24, 2026 published on AA-AgentPerf preliminary test results for Meta's newly open-weight model Muse Spark Open. The model is a 816B-parameter, 42B-active MoE, scoring 39 on the Artificial Analysis Intelligence Index. Tests show that NVIDIA B300 and GB300 each reach about 150M tokens/s per megawatt of power under throughput-first configurations; GB300 reaches about 108M and B300 about 92M at 100 tok/s output speed; AMD MI355X reaches 70M tokens/s per megawatt at 65 tok/s. Of the model's 62 layers, 46 use 2,048-token sliding window attention, reducing KV cache requirements at 1M context by about 75% versus full attention. Meta released an NVFP4 checkpoint, and AMD released an MXFP4 checkpoint. There are currently no subsequent development facts.

Artificial Analysis launches AA-AgentPerf results for Muse Spark Open

Artificial Analysis on October 24, 2026 published on AA-AgentPerf preliminary test results for Meta's newly open-weight model Muse Spark Open. The model is a 816B-parameter, 42B-active MoE, scoring 39 on the Artificial Analysis Intelligence Index. Tests show that NVIDIA B300 and GB300 each reach about 150M tokens/s per megawatt of power under throughput-first configurations; GB300 reaches about 108M and B300 about 92M at 100 tok/s output speed; AMD MI355X reaches 70M tokens/s per megawatt at 65 tok/s. Of the model's 62 layers, 46 use 2,048-token sliding window attention, reducing KV cache requirements at 1M context by about 75% versus full attention. Meta released an NVFP4 checkpoint, and AMD released an MXFP4 checkpoint. There are currently no subsequent development facts.

View story:Artificial Analysis publishes AA-AgentPerf results for Muse Spark Open

Aerial offers Agent-based automatic deployment to replace the TV's default launcher

Aerial is a tool that can replace the default launcher on Chinese smart TVs without root, and its deployment is completed automatically with the help of an Agent. According to an October 22, 2026 report by Huxiu, users connect the TV and computer to the same WiFi, enable ADB and find the TV's IP address, then copy the prompt from the Aerial official website and hand it to an Agent such as Doubao or WorkBuddy to execute; they only need to allow USB debugging in the pop-up on the TV, and the whole process takes about five minutes. The tool uses a "launcher guard" to cover 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, but boot ads still cannot be blocked. The report says that everything from the code and APK to the official website was produced through AI Vibe Coding; it gained more than 200 stars on GitHub within one day of launch, and subsequently added features such as icon hiding and adding signal input sources to the home page, and was adapted to multiple TV models.

Artificial Analysis releases Coding Agent Index v2.0

On October 23, 2026, Artificial Analysis released Coding Agent Index v2.0, replacing DeepSWE with Cognition's FrontierCode v1.1 and tightening the repository question-answering scoring; the leaderboard evaluates model-plus-harness combinations. Reports that day showed Claude Code paired with Claude Opus 5.5 (max) topping the list with 61.7 points, Claude Code paired with Sonnet 5.5 second with 59.3 points, Muse Code paired with Muse Wonder Full at 58.9 points, and the first OpenAI combination, Codex paired with GPT-6 Astra, at 54.7 points. Regarding comparisons between the old and new leaderboards, earlier reports said the average score of 21 configurations fell from 58.8% to 57.1%; later reports said the mean for the same batch of 21 configurations dropped from 54.8 to 47.4, with 20 declining, and stated that the main reason was FrontierCode's mean of about 45.1%, far below DeepSWE's 65.6%.

Artificial Analysis releases Coding Agent Index v2.0

On October 23, 2026, Artificial Analysis released Coding Agent Index v2.0, replacing DeepSWE with Cognition's FrontierCode v1.1 and tightening the repository question-answering scoring; the leaderboard evaluates model-plus-harness combinations. Reports that day showed Claude Code paired with Claude Opus 5.5 (max) topping the list with 61.7 points, Claude Code paired with Sonnet 5.5 second with 59.3 points, Muse Code paired with Muse Wonder Full at 58.9 points, and the first OpenAI combination, Codex paired with GPT-6 Astra, at 54.7 points. Regarding comparisons between the old and new leaderboards, earlier reports said the average score of 21 configurations fell from 58.8% to 57.1%; later reports said the mean for the same batch of 21 configurations dropped from 54.8 to 47.4, with 20 declining, and stated that the main reason was FrontierCode's mean of about 45.1%, far below DeepSWE's 65.6%.

View story:Artificial Analysis releases Coding Agent Index v2.0