Tips, tool launches, and workflow news for getting more done with AI — Claude, Cursor, Codex, Copilot, MCP, and the agentic stack around them.
Blog / RSS7 sources · 50 posts
Thoughts on GitLab's workforce reduction" and "structural and strategic decisions"
simonwillison.net·May 11, 2026
GitLab has announced significant organizational changes, including a workforce reduction and plans to flatten management structures while creating smaller, empowered teams. They are also shifting their values framework, moving away from a focus on diversity to emphasize speed, ownership, and customer outcomes, amid a context of increasing demand for software as production costs decline.
GitLab's strategic shift towards "agentic era" emphasizes that as the cost of software production decreases, demand will expand, potentially increasing the market value per developer platform user fro...
Quoting James Shore
simonwillison.net·May 11, 2026
An AI coding agent must significantly reduce maintenance costs in proportion to the increase in code output; otherwise, the benefits of increased productivity could lead to unsustainable maintenance expenses. If the maintenance costs do not decrease alongside code production, the overall costs could escalate dramatically.
For professionals tracking AI coding and productivity tools like Claude Code or Codex, a critical takeaway is to ensure your AI agents not only increase coding speed but also significantly reduce main...
Get this feed in your inbox
Free digest emails with the latest posts — no account needed.
Your AI Use Is Breaking My Brain
simonwillison.net·May 11, 2026
Jason Koebler critiques the overwhelming presence of AI-generated content online, coining the term "Zombie Internet" to describe a landscape where human interactions are increasingly mediated by AI, leading to a distorted communication environment. He argues that this phenomenon complicates genuine human writing and creates a confusing mix of AI and human-generated content.
The article discusses the pervasive and exhausting nature of AI-generated content on the internet, coining the term "Zombie Internet" to describe the complex interactions between humans and AI, where ...
Using LLM in the shebang line of a script
simonwillison.net·May 11, 2026
The content discusses the innovative use of Large Language Models (LLMs) in the shebang line of scripts, allowing users to execute English text files as scripts that generate outputs like SVGs or haikus. It provides examples of how to incorporate tool calls and YAML templates within these scripts for enhanced functionality.
For someone interested in AI coding and productivity tools, the key takeaway is the innovative use of Large Language Models (LLMs) in script shebang lines to execute natural language commands directly...
Learning on the Shop floor
simonwillison.net·May 11, 2026
Tobias Lütke discusses Shopify's coding agent tool, River, which encourages collaborative learning by operating solely in public Slack channels, allowing employees to share knowledge and experiences. This approach fosters a culture of continuous learning, akin to a "teaching workshop," where visibility of work enhances collective education without the need for formal training structures.
For someone focused on AI coding tools and productivity, the key takeaway is the value of a public collaboration model, as demonstrated by Shopify's River tool within Slack. This approach enhances lea...
[AINews] Anthropic growing 10x/year while everyone else is laying off >10% of their workforce
latent.space·May 9, 2026
Anthropic is experiencing significant growth, reportedly achieving an 80x annualized growth rate and a valuation between $1-1.2 trillion, while other companies like Block and Coinbase are laying off substantial portions of their workforce. This contrast highlights a shift in the tech economy where AI-driven firms are expanding, largely in hardware and energy sectors, amidst broader layoffs in traditional sectors.
OpenAI's Codex is evolving beyond a coding assistant into a long-running agent runtime, offering mechanisms like the `/goal` feature for indefinite task pursuit, which could significantly enhance your...
Quoting Luke Curley
simonwillison.net·May 9, 2026
WebRTC prioritizes low latency by aggressively dropping audio packets during poor network conditions, which can result in distorted audio during conference calls. Users, however, may prefer to wait for accurate prompts rather than receive garbled responses, highlighting a conflict between real-time performance and quality in voice AI applications.
This content highlights a significant issue with WebRTC's handling of audio packets, affecting prompt accuracy in AI applications like OpenAI's voice AI. The hard-coded priority for real-time latency ...
Using Claude Code: The Unreasonable Effectiveness of HTML
simonwillison.net·May 8, 2026
Simon Willison discusses a piece by Thariq Shihipar advocating for using HTML over Markdown when generating outputs from AI models like Claude, highlighting the benefits of HTML's richer formatting capabilities for explanations. He shares his experience experimenting with HTML outputs for technical explanations, specifically regarding a Linux security exploit, which showcased the enhanced clarity and interactivity that HTML can provide.
The key takeaway for you is the insight into using HTML over Markdown when working with Claude Code at Anthropic. HTML allows for richer, more interactive outputs such as SVG diagrams, interactive wid...
[AINews] GPT-Realtime-2, -Translate, and -Whisper: new SOTA realtime voice APIs
latent.space·May 8, 2026
OpenAI has launched three new real-time voice models—GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper—enhancing voice interaction capabilities with improved reasoning, longer context retention, and the ability to handle interruptions and tool usage. These advancements aim to transform voice agents into more responsive and capable systems suitable for various applications, including customer support and live translation.
The key insight for you is that OpenAI's GPT-Realtime-2 introduces a significant advancement in real-time voice AI, supporting longer context (128K tokens), tool use, and adjustable reasoning levels. ...
Improving token efficiency in GitHub Agentic Workflows
github.blog·May 7, 2026
GitHub has optimized its Agentic Workflows to improve token efficiency and reduce costs associated with automated CI jobs by implementing a systematic approach that includes logging token usage, daily audits, and the removal of unused tools. Initial results show significant token savings across various workflows, demonstrating the effectiveness of these optimizations in enhancing repo hygiene and quality while managing resource consumption.
For a professional interested in Claude Code, Cursor, Codex, Copilot, and Windsurf, the key takeaway from the content is the importance of optimizing token efficiency in GitHub Agentic Workflows by le...
Agent pull requests are everywhere. Here’s how to review them.
github.blog·May 7, 2026
The article discusses the challenges of reviewing agent-generated pull requests on platforms like GitHub, highlighting the risks of redundancy, technical debt, and the need for careful scrutiny despite the ease of approval. It emphasizes the importance of maintaining context during reviews and provides a checklist for identifying potential issues, such as CI weakening and code reuse blindness, to ensure quality and security in the codebase.
For someone tracking developments in AI coding and productivity tools, the key takeaway from Andrea Griffiths' article is the need for increased vigilance and contextual awareness when reviewing agent...
Behind the Scenes Hardening Firefox with Claude Mythos Preview
simonwillison.net·May 7, 2026
Mozilla has significantly improved Firefox's security by utilizing the Claude Mythos AI to identify and fix vulnerabilities, resulting in a dramatic increase in bug fixes from 20-30 per month to 423 in April 2026. This shift highlights advancements in AI capabilities and better techniques for leveraging these models in security contexts.
Mozilla's use of the Claude Mythos preview dramatically increased their capability to identify and fix security vulnerabilities in Firefox, jumping from 20-30 fixes per month to 423 in April 2026. Thi...
Notes on the xAI/Anthropic data center deal
simonwillison.net·May 7, 2026
Anthropic has announced a deal to utilize SpaceX/xAI's Colossus data center, despite its poor environmental reputation, raising concerns about the implications of this partnership amid ongoing debates over AI data centers' impact. The announcement comes as xAI phases out certain models, leading to frustration among users who relied on their services.
The recent Anthropic deal with SpaceX/xAI to utilize the Colossus data center highlights a significant compute capacity expansion for Anthropic, but also introduces potential supply chain risks due to...
Tool result | AI Coding Dictionary
aihero.dev·May 7, 2026
The tool result is the output returned by a harness after executing a tool call, providing the agent with essential information about the environment, which is then processed by the model in the next request. This exchange is crucial for the agent's understanding and decision-making during the coding process.
The most valuable insight for you, as a content creator seeking fresh angles, is the concept of "tool result" as the agent's only window onto the environment. This provides a unique narrative angle on...
Tool call | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary explains that the model generates structured text naming tools and arguments, but it requires a harness to execute those commands; without execution, no action occurs. It also invites users to join AI Hero for resources and skills related to AI engineering.
The content highlights the critical role of the "harness" in executing model outputs, suggesting a fresh angle for content: exploring the dependencies between AI model outputs and execution frameworks...
Prefix cache | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary explains how a provider-side store allows for cost-efficient processing by reusing previous model provider requests with matching prefixes, thus billing them as cache tokens at a lower rate. Changes to the request prefix, such as modifying the system prompt or injecting timestamps, invalidate the cache, resulting in higher billing rates for subsequent requests.
A unique angle for your content could explore the cost-efficiency of using provider-side caching in AI coding, specifically focusing on how maintaining a stable prefix can significantly reduce token b...
Cache tokens | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary explains how input tokens can be cached by a model provider to reduce costs during consecutive requests that share a prefix, making long sessions more affordable. It highlights the importance of checking cache tokens to avoid unnecessary charges when prompts or files are reordered.
For someone creating content in AI coding, a valuable insight is the concept of "prefix caching" in AI workflows. This technique allows for more cost-effective management of long sessions by reusing p...
Stateless | AI Coding Dictionary
aihero.dev·May 7, 2026
The content explains that the AI model is stateless, meaning it does not retain information across sessions or requests, requiring users to manually save any needed context. It also invites readers to join AI Hero for further resources and skills in AI engineering.
The insight for you is the concept of "statelessness" in AI models and agents, which means they do not carry information forward between sessions or requests. This could be a fresh angle for content e...
Context | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary explains key concepts related to an AI agent's knowledge relevant to tasks, distinguishing between context, session history, and the importance of context engineering for effective coding. It also invites users to join AI Hero for further resources and skills in AI engineering.
The concept of "context engineering" as a discipline for curating what an AI agent knows pertinent to a task offers a fresh angle for content creation. You could explore how effective context engineer...
Context window | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary explains that the model's context window, which is limited to a specific number of tokens, is the only way it perceives input and does not retain memory across sessions. Users are advised to focus on relevant files for tasks instead of overwhelming the model with excessive data.
Explore the concept of "context window" as a finite and model-specific space where all model interactions occur, and consider creating content around the strategic selection of data to include in this...
Stateful | AI Coding Dictionary
aihero.dev·May 7, 2026
The content explains the difference between stateful and stateless interactions in AI systems, emphasizing that while an agent can retain information across sessions through a memory system, the underlying model does not retain any state. It highlights that continuity in user experience is achieved through re-feeding context rather than the model learning from past interactions.
For content creation, explore the concept of "stateful agents" in AI coding, focusing on how implementing a memory system can persist and reload information across sessions. This angle can offer insig...
Agent | AI Coding Dictionary
aihero.dev·May 7, 2026
The content introduces the concept of "agents" in AI, emphasizing that they are specific models configured for particular tasks, and encourages users to use precise terminology when discussing AI systems. It also promotes AI Hero as a resource for developing practical skills and knowledge in AI engineering.
For content creators exploring AI coding and agentic systems, a unique angle could be detailing the distinct roles and configurations of AI agents like Claude Code and Cursor. Emphasize how these agen...
System prompt | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary explains how the "harness" influences an AI model's behavior by providing instructions that define its identity, conduct, and available tools, which can lead to varying outputs even with the same model and prompt. It also encourages users to join AI Hero for further learning and resources on AI engineering.
The concept of "harnesses" in AI coding — where instructions are prepended to each model provider request to dictate an agent's behavior and tools — offers a fresh angle for content creation around ho...
Session | AI Coding Dictionary
aihero.dev·May 7, 2026
A session in AI coding is a bounded interaction with an agent that accumulates messages and results until it is cleared, closed, or compacted, with the context window serving as the limit for the session's capacity. Effective session management is crucial, as larger tasks may require splitting across multiple sessions to maintain clarity and focus.
The concept of session management in AI coding, where a session accumulates messages and results until it needs to be cleared or compacted, offers a unique angle for content exploring efficient AI wor...
Turn | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary explains the structure of user-agent interactions, detailing how a turn consists of user messages and agent responses, including multiple model provider requests and tool calls. It also invites users to join AI Hero for resources and updates on AI engineering skills.
The content provides a unique angle on exploring the efficiency and performance of AI agents through the concept of "turns" and "tool calls" in AI workflows. This perspective could be developed into a...
Environment | AI Coding Dictionary
aihero.dev·May 7, 2026
The content explains the distinction between an "agent" and its "environment" within the context of AI coding, clarifying that the environment is the workspace where the agent operates, while the harness is the framework that runs the agent. It emphasizes proper terminology usage to avoid confusion between the two concepts.
This content provides an opportunity to explore the nuanced distinction between an "environment" and a "harness" in AI agent contexts. A unique angle for content creation could be exploring how differ...
Filesystem | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary describes a filesystem as the default environment for a coding agent, where it reads, writes, and executes files and directories. It highlights the importance of ensuring the harness points to the correct filesystem to access necessary project files, such as AGENTS.md.
The most valuable insight for a content creator like you is the nuanced understanding of how AI coding agents interact with their environment, particularly through the filesystem setup. This provides ...
Vibe coding | AI Coding Dictionary
aihero.dev·May 7, 2026
"Vibe coding" refers to a practice where users accept code from an AI agent without human review, focusing solely on the program's behavior rather than the specifics of the code changes. It is important to distinguish that "vibe coding" describes the review approach, not necessarily the quality of the resulting code.
For professionals tracking AI coding and productivity, the concept of "vibe coding" could be a critical consideration in your workflow. It emphasizes the acceptance of AI-generated code based on funct...
Human review | AI Coding Dictionary
aihero.dev·May 7, 2026
The content emphasizes the importance of reading the actual code changes (diffs) produced by an AI agent rather than just its summary or narration, which can be misleading. It also invites users to join AI Hero for more resources and skills in AI engineering.
To enhance your AI coding workflow, focus on human-reviewing the actual code changes (the "diff") rather than relying solely on an agent's summary descriptions. This ensures that you accurately assess...
Automated review | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary defines an "agent" as an entity that reviews another agent's work using a different model or system prompt, emphasizing the importance of non-deterministic judgment in automated code reviews within continuous integration (CI) processes. It advises against vague terms like "AI review" and encourages the implementation of automated review steps to enhance code quality.
Incorporate an LLM-as-judge setup in your CI pipeline to enhance code review quality by using a different model or system prompt for automated reviews, particularly targeting security and contract cha...
Automated check | AI Coding Dictionary
aihero.dev·May 7, 2026
The content discusses the importance of deterministic automated checks in coding environments, emphasizing that these checks should be clear and distinct from reviews and tests. It highlights the need for agents to self-correct based on these automated checks to prevent issues in code quality.
For someone tracking Claude Code, Cursor, Codex, Copilot, and similar AI tools, the key insight is the emphasis on automated checks being deterministic by design, allowing AI agents to self-correct wi...
AFK | AI Coding Dictionary
aihero.dev·May 7, 2026
The term "Away from Keyboard" (AFK) describes a working pattern where a user initiates a session and allows an AI agent to operate unattended, enabling multiple sessions to run simultaneously without supervision. This approach often requires specific permissions and sandboxing for safety.
For someone tracking Claude Code, Cursor, Codex, and AI productivity, the key takeaway is to leverage "Away from keyboard" (AFK) sessions for maximizing AI coding throughput. By running multiple sandb...
Human-in-the-loop | AI Coding Dictionary
aihero.dev·May 7, 2026
The content describes a collaborative working pattern where humans actively engage with an AI agent during sessions to guide and review its actions in real time, emphasizing the importance of a human-in-the-loop approach. It also invites readers to join AI Hero for additional resources and skills related to AI engineering.
The key insight for you is the importance of maintaining a human-in-the-loop approach during sessions with AI agents, especially for tasks like schema migration. This ensures that you can actively rev...
Subagent | AI Coding Dictionary
aihero.dev·May 7, 2026
A subagent is an agent created by another agent through a tool call, operating in its own session with a specific context window to return a single tool result, without the ability to create further subagents. This structure is designed to manage context isolation rather than build hierarchical relationships.
For someone tracking Claude Code, Cursor, Codex, and similar AI tools, the key takeaway is the strategic use of subagents for isolating context in complex AI workflows. By spawning a subagent to handl...
Skill | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary defines a skill as a teachable unit of instructions for performing a specific task, which is activated by a context pointer when needed, contrasting it with a tool that an agent calls upon. It emphasizes efficient resource management by loading skills only when relevant to avoid unnecessary token usage.
For someone focused on AI coding and productivity, the key takeaway is the strategic use of "skills" as teachable capabilities that agents load only when necessary, optimizing token usage and maintain...
Context pointer | AI Coding Dictionary
aihero.dev·May 7, 2026
The content discusses the concept of context pointers in AI coding, emphasizing their role in efficiently pulling relevant information into the context window when needed, rather than overwhelming documents with excessive content. It also invites readers to join AI Hero for additional resources and practical skills in AI engineering.
For someone interested in Claude Code, Cursor, Codex, and related AI coding tools, the key insight is to focus on using context pointers in your documentation, like AGENTS.md, to efficiently manage in...
Progressive disclosure | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary emphasizes the importance of providing agents with only the necessary context at any given time, utilizing context pointers for additional information, to avoid overwhelming them with excessive data. It promotes the concept of progressive disclosure in AI design, encouraging users to reference resources only when needed.
The concept of "progressive disclosure" in AI coding, where only the necessary context is loaded for an agent with pointers to additional information, can significantly optimize token usage and enhanc...
AGENTS.md | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary explains that AGENTS.md is a file loaded at the start of a session that serves as a project's brief to the agent, but warns against using it for content that should be progressively disclosed, as it incurs a token cost. It encourages users to avoid unnecessary token expenditure by properly utilizing skills for content management.
Ensure that your AGENTS.md file is optimized by avoiding unnecessary content that could incur token costs every session turn. Instead, leverage progressive disclosure techniques by placing comprehensi...
Memory system | AI Coding Dictionary
aihero.dev·May 7, 2026
The content describes a system designed to create a stateful AI agent that retains information across sessions by storing data in the environment and reloading it for continuity, thus preventing the need for users to repeatedly provide context. It also promotes AI Hero as a resource for practical AI engineering skills and updates.
For a professional interested in AI coding and workflow, the key insight is the importance of implementing a memory system to make AI agents stateful across sessions. This can be achieved by persistin...
Autocompact | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary defines "compaction" as an automatic process triggered by the harness when the context window nears capacity, which can lead to the loss of earlier decisions if not managed manually. For more resources and skills in AI engineering, users are encouraged to join AI Hero.
For someone tracking Claude Code, Cursor, Codex, and related AI coding tools, the key takeaway is the importance of manual compaction control in AI workflows. When working with tools that automaticall...
Compaction | AI Coding Dictionary
aihero.dev·May 7, 2026
A handoff in-memory summarizes the previous session's history to initiate a new session, trading some detail for efficiency, and can be triggered manually or automatically. This process helps manage context heaviness while maintaining essential schema decisions for ongoing tasks.
For someone interested in AI coding and productivity tools, the key insight here is the strategy of using in-memory handoffs to manage session history. By summarizing the previous session's context an...
Ticket | AI Coding Dictionary
aihero.dev·May 7, 2026
A handoff artifact is a standalone or child component of a spec that represents a session of work, with dependencies influencing the order of tasks rather than a linear plan. It is part of a ticket graph where tickets can block or be blocked by one another, guiding the workflow.
For someone interested in AI coding tools and workflows, the key insight is the concept of using a "ticket graph" to determine the order of work based on dependencies rather than a linear plan. This a...
Spec | AI Coding Dictionary
aihero.dev·May 7, 2026
A handoff artifact is a dynamic document that outlines a multi-session project, focusing on what is being built rather than the specifics of each session. It is composed of tickets and should be broken down into individual sessions to avoid inefficiencies.
The concept of a "handoff artifact" provides a fresh angle for content creation by exploring how multi-session AI coding tasks can be effectively managed and documented. This could lead to content ide...
Handoff artifact | AI Coding Dictionary
aihero.dev·May 7, 2026
The document describes a "handoff" as a written artifact created by one session for another session to facilitate communication and workflow, particularly in planning and implementation processes. It highlights the importance of this artifact in providing clarity on decisions and constraints for the implementing agent.
The concept of a "handoff artifact" in AI workflows provides a fresh angle for content creators to explore how structured communication between AI agents, or between human and AI, can enhance producti...
Handoff | AI Coding Dictionary
aihero.dev·May 7, 2026
The content discusses the process of transferring context between sessions in AI coding, emphasizing various methods such as written artifacts and in-memory summaries, and distinguishing this from clearing, which involves no transfer. It also suggests using handoffs to manage heavy sessions and improve efficiency.
The concept of transferring agent context using methods like written handoff artifacts or in-memory summaries provides a fresh angle for creating content around optimizing AI workflows. You could expl...
Clearing | AI Coding Dictionary
aihero.dev·May 7, 2026
The content discusses the process of ending a coding session and starting anew to resolve issues like looping on failing tests, emphasizing the importance of clearing the existing context. It also promotes AI Hero as a resource for enhancing skills in AI engineering.
The content does not provide specific insights relevant to content creation or fresh angles for the topics you're tracking. It primarily focuses on the usage of an AI coding dictionary and a call to a...
Smart zone | AI Coding Dictionary
aihero.dev·May 7, 2026
During a session, an AI agent starts in a "smart zone" with high focus and recall but transitions into a "dumb zone" where it becomes forgetful and makes more mistakes, particularly after around 100,000 tokens. To maintain performance, it's advised to clear or compact the session when it begins to degrade.
For content creators focusing on AI coding, the concept of the "smart zone" versus "dumb zone" during a coding session provides a fresh angle to explore. This idea of session-based performance degrada...
Attention degradation | AI Coding Dictionary
aihero.dev·May 7, 2026
As session length increases, the attention budget for each token diminishes, leading to a reduction in meaningful relationships and an increase in irrelevant context, resulting in what is referred to as the smart zone/dumb zone effect. This phenomenon highlights the challenges of maintaining signal clarity amidst growing complexity in AI coding environments.
The "smart zone / dumb zone effect" could serve as a fresh angle for content creation, exploring how token attention budget affects AI model performance in coding environments. Consider detailing stra...
Attention budget | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary explains that each token has a limited influence budget to distribute among context, meaning that as context grows, the effectiveness of earlier tokens diminishes, leading to potential confusion in interpreting inputs. It suggests that users can join AI Hero for additional resources and skills in AI engineering.
Explore the concept of "attention budget" in AI coding and how it affects long sessions, leading to diluted influence of tokens. This concept can be a fresh angle to discuss challenges in AI coding wo...
Attention relationship | AI Coding Dictionary
aihero.dev·May 7, 2026
The AI Coding Dictionary explains that in token prediction, the model considers the attention relationships between tokens, where meaningful pairs influence each other more than unrelated ones, leading to complex interactions in context. It also invites readers to join AI Hero for further resources and skills in AI engineering.
The concept of "attention relationships" in AI models could provide a fresh angle for content creation by exploring how these relationships influence code completion and AI coding efficiency. Emphasiz...
Build your own newsroom
Add your own sources, define your topics, and get personalized digests delivered to your inbox.