English text is being prepared.

Cognition releases coding model SWE-2 Flash

Cognition releases the coding model SWE-2 Flash. According to the Cognition model/Devin blog, the model is trained on Thinking Machines' Inkling-Small (276B total parameters, 12B activated parameters), and in the training pipeline a mid-training stage before reinforcement learning was added for the first time. The company claims Terminal-Bench 4.0 improved from 0.9% to 20.0%, FrontierCode 2.0 from 3.2% to 24.0%, and ExploitBench from 11.3% to 34.8%. The model is now available on Devin Desktop and CLI, priced at one-fifth of the larger SWE-2 from the same family, and three times faster (200 tok/s versus 60 tok/s); its weights are not yet public, with plans to share them securely with trusted industry partners and research institutions through a program. Cognition also claims that the model surpasses all US open-weight models on major benchmarks, including models with 3 times as many parameters; within this month, Pro, Max and Teams users can use it for free in Devin Desktop and Devin CLI.

View story:Cognition releases coding model SWE-2 Flash

Goldman Sachs Restricts Non-Engineering Employees' Use of AI Coding Tools

Since October, Goldman Sachs has tightened access to AI coding tools such as Anthropic's Claude Code and Microsoft's GitHub, mainly restricting them to developers. Previously, some traders had used these tools to monitor client trading flows and model market changes and positions. According to Bloomberg, citing anonymous sources, the adjustment has upset some traders; a Goldman Sachs spokesperson said the reason is more precise monitoring of productivity and risk management, and stressed that the tools should be given to those best suited to long-term use. The restriction does not apply to Strats employees in roles at the intersection of trading, quantitative research and software development, and Goldman Sachs' in-house assistant, which integrates models from OpenAI, Anthropic and Alphabet, remains open to all employees.

Goldman Sachs restricts non-engineering employees from using AI coding tools

Since October, Goldman Sachs has tightened access to AI coding tools such as Anthropic's Claude Code and Microsoft's GitHub, mainly restricting them to developers. Previously, some traders had used these tools to monitor client trading flows and model market changes and positions. According to Bloomberg, citing anonymous sources, the adjustment has upset some traders; a Goldman Sachs spokesperson said the reason is more precise monitoring of productivity and risk management, and stressed that the tools should be given to those best suited to long-term use. The restriction does not apply to Strats employees in roles at the intersection of trading, quantitative research and software development, and Goldman Sachs' in-house assistant, which integrates models from OpenAI, Anthropic and Alphabet, remains open to all employees.

View story:Goldman Sachs Restricts Non-Engineering Employees' Use of AI Coding Tools

Guillermo Rauch proposes Vercel CI to solve the AI coding bottleneck

Vercel CEO Guillermo Rauch published a post explaining how Vercel CI solves the AI coding bottleneck, dividing it into three layers: fast and stable infrastructure, including 20 global clusters running millions of microVMs every day; Turborepo integration, which can automatically compute the dependency graph and skip or cache existing work; and Agentic CI. He said that vercel ci in the CLI is the first step; agents can now inspect, abort, trigger, and replay CI, and Vercel Agent can assist with optimization. More capabilities will be added later to give CI engineer-like judgment. The official Vercel account previously announced that Vercel CI had entered beta, speeding up CI in internal repositories by 2-6x, and that it is optimized for agents through Vercel MCP and the CLI.

Guillermo Rauch proposes Vercel CI to solve the AI coding bottleneck

Vercel CEO Guillermo Rauch published a post explaining how Vercel CI solves the AI coding bottleneck, dividing it into three layers: fast and stable infrastructure, including 20 global clusters running millions of microVMs every day; Turborepo integration, which can automatically compute the dependency graph and skip or cache existing work; and Agentic CI. He said that vercel ci in the CLI is the first step; agents can now inspect, abort, trigger, and replay CI, and Vercel Agent can assist with optimization. More capabilities will be added later to give CI engineer-like judgment. The official Vercel account previously announced that Vercel CI had entered beta, speeding up CI in internal repositories by 2-6x, and that it is optimized for agents through Vercel MCP and the CLI.

View story:Guillermo Rauch proposes Vercel CI to solve the AI coding bottleneck

Entire launches a Git hosting platform for human-machine collaborative coding

On October 28, 2026, Entire, founded by a former GitHub CEO, launched a new-generation Git hosting platform aimed at human-machine collaborative coding, redesigning code hosting, review, CI, and Agent session memory around human-machine collaboration. The platform replaces the diff-centric PR timeline with Trail; each push is scored by a review Agent in a cloud sandbox that tracks unresolved issues, calling Claude Code, Codex, and others under the hood, with sandboxes provided by Daytona and E2B. Its self-developed Git service offers its own open-source benchmark, claiming pushes are up to about 89 times faster than GitHub, and it has integrated Buildkite and Depot as CI. There is currently no information on subsequent developments.

Larsen Cundric asks developers about the entry point for using coding agents

On October 30, 2026, Larsen Cundric asked developers: in which environment do they actually interact with coding agents at present. The options he offered included the terminal, IDE, desktop app, and mobile app. There is no subsequent development or response in the available material; the question remains at the stage of soliciting developers' points of entry for use.

Larsen Cundric asks developers about their entry point for using coding agents

On October 30, 2026, Larsen Cundric asked developers: in which environment do they actually interact with coding agents at present. The options he offered included the terminal, IDE, desktop app, and mobile app. There is no subsequent development or response in the available material; the question remains at the stage of soliciting developers' points of entry for use.

View story:Larsen Cundric asks developers about the entry point for using coding agents

Thariq suggests that the coding agent frenzy is a form of avoidance

Thariq suggests that using coding agents can lead to a kind of frenzy: feeling as though you have been working the whole time and have accomplished a lot, when in fact nothing has really happened. He believes this frenzy is often a way of avoiding other things in life. This view was published on October 30, 2026, and there has been no subsequent development or official response.

Thariq suggests that coding agent fervor is a form of avoidance

Thariq suggests that using coding agents can lead to a kind of frenzy: feeling as though you have been working the whole time and have accomplished a lot, when in fact nothing has really happened. He believes this frenzy is often a way of avoiding other things in life. This view was published on October 30, 2026, and there has been no subsequent development or official response.

View story:Thariq suggests that the coding agent frenzy is a form of avoidance

Arena.ai compares coding agent user criticism across 20,869 traces

Based on 20,869 traces from Agent Arena: Code, Arena.ai compared user criticism signals of five top models from Anthropic and OpenAI against other agents. The analysis found that frontier models differ most noticeably in three dimensions: 6.9 percentage points fewer complaints about failure-fix loops and 1.9 percentage points fewer about ignoring project context; 6.4 percentage points fewer design issues and 5.8 percentage points fewer issues with producing usable results; and 5.6, 4.9, and 4.5 percentage points fewer complaints about instruction following, task completion, and breaking existing code, respectively. There are currently no follow-up developments.

Herdr launches desktop app for terminals and coding agents

Herdr has launched a desktop app for terminals and coding agents that organizes work by space and summarizes each agent's Working, Blocked, Done, and Idle status in a sidebar. After the app window is closed, terminals continue running in the core, and spaces can also be placed on remote machines for execution via SSH. The app supports Linux, macOS (Apple Silicon), and Windows x64, includes three built-in layouts—Herdr, Terminals, and Herdr 98—and allows mod panels to be written in JavaScript. There are currently no reports of further developments.

Herdr offers a terminal and coding agent desktop app

Herdr has launched a desktop app for terminals and coding agents that organizes work by space and summarizes each agent's Working, Blocked, Done, and Idle status in a sidebar. After the app window is closed, terminals continue running in the core, and spaces can also be placed on remote machines for execution via SSH. The app supports Linux, macOS (Apple Silicon), and Windows x64, includes three built-in layouts—Herdr, Terminals, and Herdr 98—and allows mod panels to be written in JavaScript. There are currently no reports of further developments.

View story:Herdr launches desktop app for terminals and coding agents

Google Opens Android Development Tools to Third-Party Coding Agents

At the Android Dev Summit, Google announced that Android Studio and the Android CLI are being opened to third-party coding agents. According to The New Stack, Bring Your Own Agent already supports more than 40 agents that follow the Agent Client Protocol, including Claude Agent, Codex, and Google's own Antigravity, and it has been made stable in Android Studio Rabbit 2. The new android preview command lets agents render Jetpack Compose previews in the terminal; Google says this completes UI tasks 77% faster and cuts token usage by up to three times, but it did not explain how this was measured. android device remote, meanwhile, gives agents access to physical devices in Google's labs. In a demo, Claude Code used LeakCanary to investigate a memory leak and then fixed and retested it on its own; Google says such investigations are nearly twice as fast and use up to 25% fewer tokens. Google also launched Android Bench 2.0 to evaluate how models and agents perform on complex Android tasks.

Google opens Android development tools to third-party coding agents

At the Android Dev Summit, Google announced that Android Studio and the Android CLI are being opened to third-party coding agents. According to The New Stack, Bring Your Own Agent already supports more than 40 agents that follow the Agent Client Protocol, including Claude Agent, Codex, and Google's own Antigravity, and it has been made stable in Android Studio Rabbit 2. The new android preview command lets agents render Jetpack Compose previews in the terminal; Google says this completes UI tasks 77% faster and cuts token usage by up to three times, but it did not explain how this was measured. android device remote, meanwhile, gives agents access to physical devices in Google's labs. In a demo, Claude Code used LeakCanary to investigate a memory leak and then fixed and retested it on its own; Google says such investigations are nearly twice as fast and use up to 25% fewer tokens. Google also launched Android Bench 2.0 to evaluate how models and agents perform on complex Android tasks.

View story:Google Opens Android Development Tools to Third-Party Coding Agents

Cognition Launches Coding Benchmark FrontierCode 2.0

On October 28, 2026, Cognition launched the coding benchmark FrontierCode 2.0 to evaluate whether the code produced by models can actually be merged; even the highest-scoring Opus 5.5 scored only 45%, showing the benchmark is far from saturated. According to the Cognition blog, compared with version 1.1, the reference solutions modify on average 67% more files and 89% more lines of code, the codebase size is 25% larger, task coverage repositories increased by 55%, and blocking scoring criteria increased by 43%; DeepSeek V4.1 Flash dropped from 43.9% to 19.3%, and Kimi K3 dropped from 44.2% to 17.7%. On the same day, Artificial Analysis released Intelligence Index v5, listing FrontierCode 2.0 as one of the newly added benchmarks, with 60% of evaluations based on private test sets or held-out answers; Claude Opus 5.5 (max) topped the list with 58 points.

Entire launches Git hosting platform for human-AI collaborative coding

On October 28, 2026, Entire, founded by a former GitHub CEO, launched a new-generation Git hosting platform aimed at human-machine collaborative coding, redesigning code hosting, review, CI, and Agent session memory around human-machine collaboration. The platform replaces the diff-centric PR timeline with Trail; each push is scored by a review Agent in a cloud sandbox that tracks unresolved issues, calling Claude Code, Codex, and others under the hood, with sandboxes provided by Daytona and E2B. Its self-developed Git service offers its own open-source benchmark, claiming pushes are up to about 89 times faster than GitHub, and it has integrated Buildkite and Depot as CI. There is currently no information on subsequent developments.

View story:Entire launches a Git hosting platform for human-machine collaborative coding

Arena.ai compares user criticism signals of frontier coding agents

Based on 20,869 traces from Agent Arena: Code, Arena.ai compared user criticism signals of five top models from Anthropic and OpenAI against other agents. The analysis found that frontier models differ most noticeably in three dimensions: 6.9 percentage points fewer complaints about failure-fix loops and 1.9 percentage points fewer about ignoring project context; 6.4 percentage points fewer design issues and 5.8 percentage points fewer issues with producing usable results; and 5.6, 4.9, and 4.5 percentage points fewer complaints about instruction following, task completion, and breaking existing code, respectively. There are currently no follow-up developments.

View story:Arena.ai compares coding agent user criticism across 20,869 traces

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."

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