A founder who built two vertical SaaS companies — one in legal, one in finance competing with Bloomberg — argues that LLMs are systematically destroying the moats that made vertical software defensible: learned interfaces, custom business logic, and data parsing infrastructure. The $1 trillion selloff in software stocks is structurally justified because these moats were largely artificial value layered on top of genuinely scarce data. What survives is narrower: proprietary data, compliance infrastructure, and workflow orchestration at scale.
The business logic your competitors spent a decade encoding in software can now be replicated in a markdown file in a week — meaning the defensibility of your vertical software investments must be urgently reassessed.
The next $1T company will be a software company masquerading as a services firm. Every founder building an AI tool is asking the same question: what happens when the next version of Claude makes my product a feature? They’re right to worry. If you sell the tool, you’re in a race against the model. B
Andrej Karpathy has coined "agentic engineering" to describe the new default in software development, where developers orchestrate AI agents rather than write code directly 99% of the time. He believes developers could be 10x more productive by properly leveraging available tools, while warning that the programming profession is being dramatically refactored. He capped the observation by describing a fully autonomous AI workflow as feeling like "post-AGI."
If one of the world's leading AI researchers says developers are now primarily agent orchestrators — not coders — your hiring profiles, team structures, and productivity benchmarks may already be out of date.
Anthropic's Claude Sonnet 4.5 is their most capable Sonnet model, leading on coding benchmarks (77.2% SWE-bench) and computer use (61.4% OSWorld). Most significantly, it sustains focused performance across 30+ hour multi-step tasks — a threshold that makes autonomous agent workflows operationally viable. Pricing stays flat at $3/$15 per million tokens.
The 30-hour sustained focus capability moves AI agents from controlled demos into genuine operational infrastructure — making it worth revisiting any AI initiative your team shelved because agents weren't reliable enough for long-horizon tasks.
Greg Isenberg outlines a 7-step framework for building AI-first SaaS in 2026, centered on mapping workflows, quantifying pain points, and doing the work manually before automating it. He predicts 100K+ SaaS layoffs will spawn a generation of 1-10 person micro-founded AI businesses. The core insight: the best AI products start as services, not software.
AI-first SaaS success in 2026 will belong to founders who do the workflow by hand first — domain depth and judgment, not design or funding, are the new defensible moats.
A maintainer attempted to relicense the chardet library by rewriting it with Claude Code, but Simon Willison argues this fails the legal standard for a clean-room rewrite on two counts: the maintainer carries 12 years of codebase knowledge, and Claude's training data likely included the original code. This creates an unresolved legal question about whether AI-assisted rewrites can ever produce genuinely clean IP.
Assuming AI-generated code is free of IP entanglement is legally untested — companies using AI to rewrite or reimplement existing software may be inheriting liability, not escaping it.
Kieran Klaassen's compound engineering system — 27 agents, 21 commands, 14 skills — demonstrated that matching AI model speed to task type can deliver 5x more design iterations in the same timeframe. Separately, upgrading email classification from Gemini Flash 2.0 to 2.5 improved accuracy, summary quality, and reliability in production. The core lesson: the best AI architecture routes different tasks to different models, not the same powerful model to everything.
Compound AI systems that match model speed to task type can multiply creative output 5x — making model selection strategy as important as model capability for leaders building AI-powered operations.
Hamel Husain has open-sourced a set of 'eval skills' for coding agents, drawn from observations across 50+ companies and 4,000+ students, to help teams systematically audit and improve their AI evaluation setups. The centerpiece is an eval-audit skill that runs diagnostic checks across six areas and surfaces prioritized problems. As coding agents increasingly run experiments and instrument production systems, the quality of your eval infrastructure directly determines how reliably and safely you can deploy AI.
Teams deploying AI agents without a rigorous evaluation framework have no reliable signal on whether their systems are improving or degrading — this open-source resource gives leaders a proven, structured starting point to close that gap.
Dan Shipper predicts that by 2026, most new software will be a thin UI layer over general AI agents — features become prompts, and every SaaS product either becomes an agent platform or becomes irrelevant. The so-called SaaSpocalypse has already wiped roughly $1 trillion from SaaS valuations, signaling this is a structural shift, not a correction.
If your software product's value lives in its features and workflows rather than the intelligence underneath, it is already at existential risk.
Enterprise finance teams are drowning in manual "glue work" between software systems — and AI agents, not more software, are the fix. By deploying agents to handle AP exception resolution, reconciliations, and vendor communications inside existing tools, finance teams can cut month-end close from 12 days to 5. The catch: 87% of AI pilots fail because they're treated as standalone demos rather than true workflow integrations.
The highest ROI from AI in finance comes not from replacing ERP systems but from deploying agents to handle the repetitive connective tissue between them — a targeted approach that can cut close cycles by 40% and eliminate up to $100K/month in manual invoice processing costs.
Anthropic's Claude Code is now entirely self-written — 259 PRs and 497 commits authored by Claude Code running on Opus 4.5. New commands /simplify and /batch bring parallel agents into everyday code quality and migration workflows. This marks a shift from AI-assisted coding to AI-operated software development.
When the team building the leading AI coding tool trusts it to write 100% of its own production code, the case for keeping AI at arm's length in your engineering org has effectively collapsed.
Anthropic has launched Claude Cowork, an enterprise agent platform featuring MCP connectors for tools like Google Workspace, DocuSign, FactSet, and Harvey, alongside prebuilt plugin templates for functions ranging from HR to investment banking. Enterprises can also build private plugin marketplaces connected to internal repositories. This marks Anthropic's most direct move yet into replacing or augmenting enterprise software at the workflow level.
Anthropic is shifting from AI model provider to enterprise operating layer — executives should assess whether Claude Cowork could displace or consolidate existing departmental software before their vendors do it first.
Anthropic's Claude Code Security has identified over 500 zero-day vulnerabilities in production open-source code — including bugs that survived decades of expert review — by reasoning about code rather than matching patterns. It caught a heap buffer overflow in CGIF that even 100% code coverage fuzzing missed. This marks a fundamental shift from tools like CodeQL and Semgrep, moving automated security review from pattern recognition to genuine reasoning.
If AI can now find vulnerabilities that decades of expert review and best-in-class fuzzing missed, your organization's current security posture is almost certainly based on an incomplete picture of your actual risk.
Nicolas Bustamante shares hard-won lessons from two years building Fintool, an AI agent platform achieving 97% accuracy on financial benchmarks with clients like PwC and Kennedy Capital. The core insight: regulated industries require fundamentally different architectural choices — in context management, tool design, and evaluation — not just more capable models. The gap between a compelling AI demo and a production-ready system is measurably wider in financial services.
In regulated industries, AI accuracy is an architectural problem that must be designed for from the start — companies that treat it as a fine-tuning step will find the production gap far more costly than anticipated.