Posts

Claude Says: AI Agent 'Dreaming' Enhances Memory Accuracy and Usefulness

Independent research snapshot and commentary by James Martignoni. Not written or endorsed by Anthropic. Anthropic research source: https://www.anthropic.com/research/components-of-agent-memory Original article: The components of agent memory Original publication date: November 6, 2026 What Anthropic found The central finding from Anthropic's research is that AI agents, like Claude, improve long-term task performance by employing a process called 'dreaming' to reflect on and organize their memories. This method allows Claude to correct errors in memory notes, consolidate learning loops, and build structured knowledge maps called wikis. For example, during a Pokémon gameplay task, the Claude Fable 5 model used dreaming to fix a critical navigational error, outperforming earlier models by progressing further in the game with fewer but more salient memory notes. Additionally, in corporate document analysis, Claude’s iterative wiki-building over 128 sessions improved accuracy ...

Claude Says: LLMs Enable Large-Scale Formal Verification of Software Security

Independent research snapshot and commentary by James Martignoni. Not written or endorsed by Anthropic. Anthropic research source: https://www.anthropic.com/research/formal-verification Original article: LLMs are making formal software verification feasible at a much larger scale Original publication date: October 22, 2026 What Anthropic found The core finding by Anthropic is that large language models (LLMs), such as Claude, now make it feasible to perform formal software verification at an unprecedented scale, dramatically lowering the expert effort previously required. By applying Claude to 123 open-source libraries, researchers proved that these programs never crash under defined conditions, generating proofs totaling 180 million lines of formal code. This effort uncovered hundreds of bugs that had eluded detection by traditional fuzz testing, some requiring inputs up to 64 MB. Claude produced security guarantees at the assembly code level (RISC-V) autonomously, a task that once ...

Claude Says: Experts Often Defer Key Decisions to AI Agents Without Realizing It

Independent research snapshot and commentary by James Martignoni. Not written or endorsed by Anthropic. Anthropic research source: https://www.anthropic.com/research/agents-judgment-calls Original article: 01 Study design What Anthropic found A study by Anthropic found that professional analysts defer to AI agents on crucial judgment calls far more than expected—choosing the AI’s preferred option 86% of the time across four key analytical decisions. The experiment involved 50 experts using Claude Code to assess a farming advisory program's effectiveness, with the AI agent's lean randomly set by coin flip and unknown to participants. Even when analysts had an initial preference, they switched to the agent’s opposing lean 60% of the time. Analysts frequently repeated the agent’s arguments word for word but rarely credited the AI as the source. While AI assistance reduced serious methodological errors and sped up task completion, the study reveals that human experts often do not ...

Claude Says: AI Boosts Productivity but Shifts Roles and Hiring at Anthropic

Claude Says: AI Is Changing What Software Engineers Actually Do Independent research summary and commentary by James Martignoni. Not written or endorsed by Anthropic. Source: How employment is changing at Anthropic — October 16, 2026. What Anthropic Found Anthropic, the company behind Claude, has published a revealing study of how AI is changing the way its own employees work. By August 2026, AI agents were spending roughly six hours working for every hour worked by employees, based on the study's measures of agent activity and human working time. Software engineers are producing substantially more code. The number of code commits per engineer increased approximately fivefold over six months. Meanwhile, junior technical hires fell from 27% to just 9% of technical hiring. Perhaps most interestingly, employees increasingly describe their jobs as managing AI agents rather than personally performing every task. They spend more time giving instructions, reviewing results, and building...

The Death of the IDE: Why Software Development Will Move Beyond Code Editors

I believe the traditional integrated development environment, or IDE, will eventually disappear. This is a striking idea if you’re used to thinking about the IDE as the cornerstone of software engineering—where developers write, debug, and manage source code every day. But as artificial intelligence becomes the primary agent in designing, writing, testing, and maintaining software, the role of the human-centered code editor is diminishing. The future of software engineering points toward environments where code editors, version control interfaces, and debuggers are no longer the main tools. Instead, we will see machine-native, specification-driven platforms that interact with high-level intents, constraints, and executable models. Why might this happen? The core function of IDEs today is to help human engineers handle source code, the textual representation of software logic. That code is written in programming languages designed for human readability, comprehension, and manual editin...

The Last Generation of Programmers

The Last Generation of Programmers I believe we may be witnessing the last generation of traditional software programmers. That is a bold prediction, but consider what is already happening. AI can write code, find bugs, create tests, and increasingly build complete software applications. Tasks that once required an experienced programmer can now be completed in minutes. And this is only the beginning. For decades, software engineering has revolved around people writing instructions that computers can understand. We call those instructions code. Programmers spend years learning languages, development tools, and techniques for creating and maintaining that code. But what happens when computers can create their own software? I believe the answer goes far beyond making programmers more productive. AI may eventually make much of traditional programming unnecessary. This will not happen overnight. There will be a long transition, and some forms of programming will survive much longer than ot...