Paul Dix

Another development in the seemingly never-ending cacophony of conflicting messages for software engineers over the last three years. Engineers are "more important than ever" at Anthropic despite the fact Claude is doing a lot of the legwork with coding these days. Yet CEO Dario Amodei thinks AI will be doing "most, maybe all, of what software engineers do end-to-end" within six to 12 months. The script will likely flip next month.

7

1 Comment

Avoiding Claude Code because the terminal feels intimidating? Cowork changes that. Same power, zero command line. Anthropic released Cowork today and I've been testing it. Here's what makes it different: Cowork lets Claude interact with your computer through a visual interface. No typing commands. No scary black screens. What you can actually do with it: - Organize messy file folders automatically - Turn a pile of receipts into a clean spreadsheet - Draft reports by pulling data from multiple documents - Rename hundreds of files based on patterns - Create presentations from raw research notes This is Claude Code's capability through a visual interface. The same parallel processing that developers use, but accessible to non-technical users through Claude desktop. Worth exploring if you spend hours on repetitive computer work.

3

This extension of the 2025 success of Cursor and Claude into more general purpose work were on the prediction lists for 2026. We have not even hit mid January and Anthropic has just released Cowork as a preview. Cowork can take on many of the same tasks that Claude Code can handle, but in a more approachable form for non-coding tasks. Across 2025, LinkedIn was a great place to watch all the ways people have been adapting Claude Code for their daily work in product management and design. Anthropic's preview makes it more accessible outside the tech industry - media for example. These "Cursor-style" LLM apps represents a new layer of software that bundles and orchestrates language model capabilities for specific vertical industries. These apps act as a "thick" layer that organise generally capable models into "deployed professionals" by supplying them with the necessary tools and data. They will have four core components: 1. Context Engineering: The process of supplying the LLM with private data and feedback loops specific to a vertical. General-purpose models do not inherently know an organisation's "tribal knowledge" or specific history, so context engineering differentiates a company’s AI from its competitors and removes barriers to AI adoption. It also requires consolidating fragmented information— Slack threads, strategy docs, and customer records—into a format the agent can use to perform specialised tasks. 2. LLM Call Orchestration: These apps orchestrate multiple LLM calls behind the scenes. Rather than relying on a single prompt, the app manages "infinite minds" or teams of agents, queuing tasks to run autonomously over extended time horizons. 3. Application-Specific GUI: A "Cursor/Claude-style" app provides a native visual interface tailored for a human in the loop, moving beyond the traditional "command console" chat interface. 4. The "Autonomy Slider": This component allows the user to control the degree of independence the AI has, effectively shifting the human’s role from a "worker" to a "manager of agents". As AI capability moves toward "agentic autonomy," the slider enables a user to delegate long-horizon tasks while maintaining oversight. This helps move the human out of the immediate workflow loop and into a supervisory directors position, where they provide the final judgment and "escalation paths" for the AI.

2

Here's an interesting one -- Sourcegraph is spinning out its AI coding agent Amp Code as a standalone business, with Sourcegraph CEO and co-founder Quinn Slack taking the reins at the new company. Main takeaway: AI-native coding tools are big enough and strategically important enough to warrant their own companies, especially when it's already profitable (as Amp says it is). Full story here: https://lnkd.in/eGP3n8bN

8

Running evals on the new Claude 4.5 Haiku this morning—seemed like a good time to share what we’ve built for evals at Boom. The core implementation: an automated evaluation pipeline that runs every supported model against an ever growing set of test cases for our critical path inference operations. Quality, latency, and cost benchmarked in real time against production workloads. But here’s what this actually delivers for our customers: **Guaranteed uptime, not just promised uptime.** When Anthropic or Google has an outage, our system automatically routes to the next best model. Our customers’ workflows don’t stop. They often don’t even notice. We’ve decoupled their reliability from any single provider’s reliability. **Immediate access to model improvements.** Claude 4.5 Haiku dropped last week. By this morning when I finally got a few minutes to add it to our model directory, I had complete performance data across our stack. If it improves customer outcomes, we can integrate it today. Our customers get the benefit of frontier model advances within hours, not months. **Protection from silent degradation.** The harder problem isn’t when services go down—it’s when they stay up but quality degrades. Our eval system monitors this continuously. If we detect accuracy dropping on a live service, we route traffic away before it impacts customer results. The chart shows what I’m looking at right now: Haiku absolutely crushing latency on most of our critical path operations (e.g. 11s vs 26-64s for other models) while maintaining 99%+ accuracy. That’s a customer experience improvement we can ship immediately. The power is in having evals that actually matter. Test cases from realistic production scenarios. Metrics that map to customer outcomes. When you can run any model against your ground truth in minutes, you stop guessing and start knowing. Model selection becomes an empirical question, not an intuitive or philosophical one. Your customers care about outcomes, not which model you’re using.

27

2 Comments

Some real progress from Anthropic. https://lnkd.in/e64vKRwc "With Programmatic Tool Calling: Instead of each tool result returning to Claude, Claude writes a Python script that orchestrates the entire workflow. The script runs in the Code Execution tool (a sandboxed environment), pausing when it needs results from your tools. When you return tool results via the API, they're processed by the script rather than consumed by the model. The script continues executing, and Claude only sees the final output." One problem with older tool-calling approaches, like MCP, is that the agent is responsible both for orchestration and execution. Unimportant work-in-progress can quickly overwhelm an agent's context window. Eliminating the need for agents to handle execution allows them to focus on the orchestration tasks they do best. Think of your agents as programmers, not programs.

9

A new paper from SJTU and GAIR (daVinci-Env) makes a compelling case that we've been looking at the wrong variable in SWE agent training. The premise is refreshingly honest: the open community doesn't lack clever RL algorithms or agent architectures. What it lacks is a large-scale supply of executable, verifiable training environments. So they built one. 45,320 Docker environments across 12.8k Python repositories, synthesized from 572K GitHub PRs through a multi-agent builder pipeline. The environments come with Dockerfiles, evaluation scripts, and distributed construction infra - all open-sourced. The engineering decisions matter more than the headline numbers: -> Builder decomposed into retrieval / setup / evaluation / analysis roles (not one monolithic agent) -> Dual-run validation protocol: test-only must fail, test-with-fix must pass -> Structured failure routing that becomes free process supervision for training -> Success-frequency curriculum: keep instances where the model succeeds 1-2 out of 4 attempts The most convincing result is the environment-source ablation: holding model and scaffold constant, swapping only the training data source, OpenSWE beats SWE-rebench in all four settings - up to +12.2 points. That's direct evidence that environment quality is a first-order training variable. Where the paper is weaker: the "difficulty-aware curation" story lacks clean ablation, the 32B headline slightly oversells relative to the strongest RL baseline (+1.0, not +4.6), and whether the curated trajectories are fully released remains unclear. The real takeaway for anyone building post-training pipelines: verifiable feedback loops and environment quality may matter more than the choice of optimization algorithm. This paper has quietly assembled every ingredient needed for an RL-for-code environment engine - real repos, executable sandboxes, verifiable outcomes, structured error attribution. The next step is online policy improvement on top of that substrate. https://lnkd.in/e2cFM-zu

5

We did not build AlturaQuantera app for the DORA deadline. Deadlines are loud. Real work is quiet. When DORA hit, a lot of teams did what they had to do: scramble, document, align just enough to pass the first assessment. External consultants, war rooms, late-night evidence hunts. It worked, at least on paper. What I kept seeing after the rush was different: - controls that were implemented but not really owned - policies copied from templates no one had read twice - audits treated like single events instead of recurring checkpoints - compliance sitting in SharePoint, reality sitting in production That gap between "we did DORA" and "we live with DORA" is where most of the pain lives. Not in the law itself, but in the operational residue of doing it as a project instead of a system. AlturaQuantera came out of that pain. It is not a get compliant fast tool. It is a workspace for the months and years after the consultants leave, when systems change, vendors rotate, incidents happen, and someone has to explain how all of that still fits DORA. It is shaped for reassessments that do not start from zero. For evidence that ages alongside your architecture. For policies that evolve instead of accumulating as versions in a folder called Final. Most teams already survived the deadline. The real question now is quieter, and harder: will your compliance survive your normal operations? #DORA #compliance #SaaS #fintech #riskmanagement

5

A new era in the AI race came slowly, then all at once. Software engineers and tech first adopters have been raving for months about the capabilities of Anthropic’s Claude for coding. Some described how it completely revolutionized their work, leaving them in a state that alternated between awe and existential dread. Many also wondered: How long before it goes beyond software engineering? This week offered a resounding answer to that question after Anthropic released a handful of what it called “knowledge work plugins” on Github, a coding site. Far from end-to-end software solutions, the tools offered a whiff of possibility for what might lie in store across a whole range of job categories, including legal, sales, marketing, product management and more. The mere prospect of disruption for software companies tanked the stock market, leading to one of the biggest selloffs in years. Then, Anthropic released Super Bowl ads that took aim at rival OpenAI’s decisions to introduce advertising into ChatGPT (ads that OpenAI Chief Executive Sam Altman said were misleading and disappointing). Finally, Anthropic put out the latest model of Claude, Opus 4.6, which included coding improvements and other functions that appear to take further aim at knowledge work. The market rebounded Friday, but the episode showed how Anthropic has fought its way to the forefront of the AI race. Read the full story here on Anthropic’s wild week: https://lnkd.in/gXCiEjwT

35

5 Comments

In today’s installment of the Plan of Record Substack series, I break down the most common symptoms of prioritization failure in SaaS R&D orgs — from overcommitment and hidden work to heroic culture and black-box delivery — and begin to unpack the root causes behind the dysfunction. Here’s the link: https://lnkd.in/e6fdcNNj #SaaS #ProductManagement #EngineeringLeadership #R&D #Prioritization #Execution #CPO #CTO #CPTO

23

I'm not a hype person, I'm actually pretty cynical about software. (Who needs another app, right?) That being said, Claude Code is a big deal. A big part of what I do professionally is communicate technical topics in an accessible, practical way. I've found myself struggling to communicate why Claude Code is having such a large impact for people that try it. We talked about this yesterday in our training office hours with the folks from our December class. Someone shared this quote from an article: "But today, at least, Claude Code is the most important piece of AI technology on the market because it delivers on the core promise of AI: Dramatic acceleration of human potential and a contemporaneous democratization of opportunity." (https://lnkd.in/d8ykjG7k) And then today Ethan Mollick shared his thoughts. Ethan has consistently been one of my favorite voices in AI. He's been following things closely from the day that ChatGPT was launched. I actually hate getting on the internet over and over again and saying try Claude Code, but I'm going to keep doing it because it's the most important thing you can do today to understand what's possible. Let me say it a different way, if you haven't tried Claude Code, you have NO IDEA what AI is capable of. Take 30 minutes and try it. https://lnkd.in/dRzgtTk3

20

3 Comments

This week, Anthropic's Claude 4 seems to be again on a roller coaster: after pushing Amazon's new Kiro IDE (which was briefly available for download before an avalanche of demand slammed it shut), his coding performance plummeted noticeably. Tool usage has dropped. The reason? Degraded. And as usual at Anthropic, no one knows why. We just get silence and a fresh dose of usage limits. Meanwhile, Microsoft's Copilot is getting a makeover literally. The new "appearance" feature could soon allow it to simulate traits like age or personality. Whether it's charming or dystopian depends on your tolerance for digital companions with synthetic charm. But one thing is clear: AI is no longer just smart, it's present. Speaking of IAG, Meta appointed Shengjia Zhao to lead its new Superintelligence Lab. If that name doesn't ring a bell, it soon will: Meta is delving into fundamental models with AI ambitions, indicating that the "big three" (Meta, OpenAI, Google) are gearing up for the final showdown. But it's not all polished demos and ambitious projects. A troubling moment at Sketch.dev exposed the dark side of AI programming assistants: a subtle AI-written database optimization caused a crash in production under load. Reminder: just because your AI can write code doesn't mean it should. Test, test, and test again. Mistral's Codestral also launched, quietly stealing the show. With 22 billion parameters and support for over 80 languages, it's open-source, meaning it has no barriers. Elsewhere, Google's NotebookLM received a major update powered by Gemini. It now summarizes YouTube videos and drafts of content like your personal caffeine research companion. Think of it as a fusion of CliffsNotes and ChatGPT: optimized, hyper-efficient, and incredibly good at synthesizing ideas. In the worst-case scenario for cybersecurity, researchers have revealed AI-powered malware that thinks for itself. It interprets its environment, adapts, and executes evasive strategies with chilling precision. It's not HAL yet, but it's time we reconsider the true meaning of "autonomous threat." Finally, some news that caught my attention: A new concept unlike the transformers and tokens we know today, the Hierarchical Reasoning Model (HRM) concept promises to make the large models we know more efficient. In its first milestone, a small model of just 27M was able to beat o3-mini on pre-trained tasks. for more information follow me here: https://lnkd.in/gNVYeZST

4

In this presentation, Colin cogently articulates the rationale for choosing Components over Containers and why Agentic Systems provide a catalyst for that paradigm shift (especially from around the 19 minute mark).

2

You're putting out articles, but AI systems are just scraping them into data chunks. They treat your work like a library subset, where the metadata often matters more than the actual writing. No bueno. This disconnect is damaging how content gets licensed and valued. Last time I wrote for Creative Licensing International (CLI), my piece "The Publishing Industry’s Napster Moment" ended up as their best performer so far. Quite proud of that, tbh. As a result, they asked me to do another one. In this latest article for CLI's Content Licensing Brief, "The Commodity Paradox: Why AI Eats Tokens, Not Articles," I dig into the key challenge for publishers and AI labs: why AI platforms license fragments ("chunks") instead of whole articles, and how that changes the leverage for publishers. Think of cases like the The New York Times' lawsuit against OpenAI leading to settlements, or the recent Anthropic agreement to payout $1.5 billion to authors in September 2025, yet publishers still scramble for fair leverage. Even relatively large deals, like Amazon paying the NYT around $20-25 million a year, often feel like scraps compared to the tech giants' gains. If content licensing, AI deals, or rights management is your world (and you're over the vague AI hype), I wrote this thinking of folks like you; practical breakdown, no fluff. Read here: https://lnkd.in/eA3wExuG P.S. If this is your day job, subscribe to CLI's Brief; they convert your IP into a real asset. Things are changing fast, don't get caught out.

26

1 Comment

Earlier this month Open Source Collective repurposed ~$27k of stranded project funds to support the foundation on which those projects were built. We used https://funds.ecosyste.ms/ to distribute money to the most critical components within several ecosystems, sending money to hundreds of maintainers. Details at https://lnkd.in/eyJSDea2

36

1 Comment

I turned 3 projects worth of AI agent lessons into a single npx command. My last post about building AI development teams got a lot of attention. The number one question: "How do I actually set this up?" The honest answer was painful — it took half a day just to configure Project C. Eight agent definitions, skill files, file ownership matrices, settings... all by hand. And if Claude Code's spec changed? Your carefully crafted prompts were already outdated. So I built create-agent-system. npx create-agent-system One command. Pick a preset. Get a working agent team. What it does: - 3 presets (solo-dev / small-team / full-team) — opinionated defaults, not blank canvases - 8 agent types with role isolation and file ownership - 8 skill packages (scoring, TDD, visual QA, code review, ADR writing...) - Intersection skill computation — agents only get skills relevant to their preset - /sync-spec skill that validates your setup against the latest official docs The key design decision: scaffold once, stay up to date forever. Most scaffolders generate boilerplate and walk away. create-agent-system bundles a sync-spec skill that checks your configuration against  the latest Claude Code documentation via Context7 MCP. Your agents evolve with the platform. The technical highlight I'm most proud of: intersection skill computation. Each agent has default skills. Each preset activates certain skills. The agent gets the intersection — not the union. Why? Because irrelevant skills pollute the agent's context. I learned this the hard way when a Backend agent started demanding screenshots for visual QA checks. 5 things I learned building this: 1. Good defaults are the best documentation 2. Presets are opinions — opinionless tools are unusable 3. Scaffolding without sync is half the job 4. More agents = more important to filter skills precisely 5. There's a right level of abstraction — 100-line Handlebars templates are not it It's open source: https://lnkd.in/gmHwBYfA Full writeup with design philosophy, preset comparison, and code examples: https://lnkd.in/g62cPXCX #ClaudeCode #AI #OpenSource #DeveloperExperience #AIEngineering

4

ccusage gave me my first real window into Claude Code, mostly the cost angle, how many tokens a session burns, how that scales across a week of work. Useful. But the more I looked at it, the more I realized cost is just one dimension. I still have almost no visibility into what Claude Code actually does during a session, which tools it calls, how many turns a task takes, where it gets stuck before finding a solution. That's the gap Langfuse is built to close. Langfuse is an open-source LLM engineering platform that gives you traces, metrics, prompt management, and evaluation tooling for AI applications. It was acquired by ClickHouse earlier this year, which tells you something about where the LLM observability space is heading. The self-hosted option is what caught my attention, all the observability data stays local, which matters when you're tracing sessions that include your own code and prompts. It has a native Claude Code integration via the hooks system, and I want to see what that actually surfaces in practice. Does seeing tool usage patterns change how I prompt? Are there session structures that correlate with better outcomes? Can I hook this up alongside other agentic tools to compare how they behave under similar tasks? The integration is still relatively new and the community is working out some rough edges around OpenTelemetry compatibility. That's part of what makes it interesting to explore. If you've already set this up, I'd be curious what you've learned from it. Link: https://langfuse.com 74/365 #AI #Observability

2