6 min read · 13 sources

The biggest story today isn’t a product launch. It’s a playbook for buying businesses and rebuilding them with AI agents. McKinsey puts the number at roughly $5 trillion of American businesses changing hands by 2035, most of it boomer-owned firms with no succession plan. The thesis: buy the customers, licenses, and trust, then replace the delivery model with agents. An accounting firm in the example processed 7,000 tax returns with AI, saving accountants 31% of their time and collapsing one 180-hour annual task to 15 hours. A services firm running at 5-10% EBITDA margins gets rebuilt to run at 30-40%, tripling or quadrupling profit on the same revenue.

This is the concrete deployment pattern engineers have been waiting for. It’s not another chatbot wrapper - it’s a structural arbitrage on labor costs, applied to businesses that already have revenue and distribution. If you’ve been wondering where agent economics actually land, this is the math.

A $5 million revenue company can be a better acquisition target than a $50 million one, because AI has cut the time to build internally.

The Neo Neo Lab: Why Enterprises Are Training Their Own Models Now

Source: aspiringforintelligence.substack.com ↗

Cheap, abundant base intelligence has made building your own models more attractive, not less. The argument in The Rise of the Neo Neo Lab is that the next generation of enterprise software is built around proprietary intelligence, not proprietary data and workflows.

The evidence is concrete. Thomson Reuters shipped a model named “Thomson” that performed competitively with frontier models like Claude Opus 4.8 and ahead of GPT-5.5, using less than 10% of their content for training. Morgan & Morgan committed $1 billion to AI over a decade after spending $300 million since 2021 on its in-house MX2 platform. Harvey, the legal AI tool, is transitioning from an app to a model with “Tenet,” a post-trained open-weight model based on Kimi K3.

For engineers this is a shift in default architecture. Instead of relying on API-based frontier models, enterprises are post-training open-weight models on domain-specific data. The frontier labs push general intelligence; enterprises teach it what only they know. If you have the proprietary data and workflow volume, you have the raw materials.

The Irreplaceables: The Employees You Never Let Go

Source: saastr.com ↗

The framework from SaaStr is simple: an “Irreplaceable” is someone you’d find budget for within 48 hours, before knowing what they’d do. That distinguishes them from top performers, who you’d rehire into an existing role.

Three traits define them: they take work off your plate without being asked and finish it; their output doesn’t change when nobody is watching; and they bring you bad news early. The warning is that this usually happens at scale - new executives don’t think these employees are so irreplaceable, and they’re wrong. As skill sets go stale in about 18 months, the people who know how and why everything works are exactly the ones you need to figure out the next move.

Gusto's Cofounder: AI Made Your Revenue Worth Less

Source: x.com ↗

Gusto’s cofounder explains how AI changed their M&A strategy, using the acquisition of Guideline, a 401(k) provider, as the example. The insight: AI lowers the cost of building things internally, so the bar for acquiring a company is now about “acceleration value” to Gusto, not raw revenue.

The math is stark. If buying a company saves 18 months of internal development worth $50 million to Gusto, a business with $5 million in revenue might fit at a 10x multiple. One with $50 million in revenue probably won’t - your revenue raises your asking price without increasing how much development time the buyer saves. An acquirer may love your product and still decide it’s faster to build its own.

For founders, this reframes the exit conversation. Your revenue multiple is no longer the metric that matters; your team’s speed and the uniqueness of your codebase are.

n8n Agents: From Fixed Workflows to Self-Planning Steps

Source: blog.n8n.io ↗

n8n has introduced Agents, a feature that lets users describe a goal in plain language, give it a model and tools, and let it figure out the steps itself. Agents can be reached via Slack, run on a schedule, or be called from any workflow, and they can use existing workflows as tools.

The existing AI Agent node is unchanged; a new “Message an Agent” node lets workflows call agents for open-ended steps. This matters for engineers because it handles back-and-forth, open-ended tasks that are hard to model as fixed sequences. The same agent runs in every place it’s called, which is the key architectural point.

Parallel and Alexandria: Data Connectors for Agents

Source: parallel.ai ↗

Parallel’s data connectors are integrations for various data sources to be used in workflows. It lists “Allium” for standardized blockchain data, and “Parallel Search Fast” as a fast, cheap web search API for agents. Several connectors are listed as “Coming soon,” including company/professional contact data, business identity/credit data, customer behavior predictions, and prediction-market prices.

Firecrawl Alexandria is a data library giving AI agents access to sources web search can miss: live web, official providers, and specialized indexes. It lists 93 providers, 640 capabilities, 20 categories, and 113M+ indexed sources, including government records, product catalogs, code/docs, and financial data. It connects via MCP or CLI with npx -y firecrawl-cli@latest init --all --browser.

For engineers, these are the plumbing for giving agents structured, licensed data instead of scraping the open web. The quality of agent answers is bounded by the data they can reach.

The Cautious Tale of Cal AI: Luck, Not a Template

Source: operatorsnotebook.beehiiv.com ↗

The founding story of Cal AI went from nothing to #1 in Health & Fitness and $50M ARR in eighteen months before selling. The author’s warning: the team’s success was largely due to luck, not a repeatable process. Before the revenue, one founder’s mom had to co-sign the company bank account, and two of the four founders were still in high school.

The chain of events that led to meeting his co-founders involved a cousin’s YouTube recommendation, a cold DM, and meeting a future cofounder while working at another startup. There was no recruiting process to copy. The advice is more realistic: look for someone who is available and can learn, and expect to offer enough money or equity to make joining worthwhile. Hard work and intelligence only increase your “surface area for luck.”

The Value Accrual Flywheel and Supervoting

Source: konvoy.beehiiv.com ↗

The Value Accrual Flywheel explains how software accrues value by creating dependencies that compound over time and raise switching costs. These dependencies compound internally as the product embeds deeper into workflows, and externally through network effects. Dependencies inside critical systems get a value multiplier because ripping them out threatens the business itself. The resulting revenue funds new features that create new dependencies - a flywheel. This is why certain products become “systems of record,” and AI is reshuffling which dependencies matter.

On the governance side, supervoting separates ownership from control. Figma’s co-founder Dylan Field held ~9% of equity but over half the voting power through Class B shares (15 votes per share) and an irrevocable proxy from his co-founder, giving him veto power over the $20 billion Adobe merger. Supervoting is not uncommon in VC-backed companies and is distinct from non-voting shares. For founders, it’s a mechanism to maintain control even with a small economic stake.

What Would a Serious AI Product Look Like?

Source: blog.glyph.im ↗

The critique from Glyph is that current AI products don’t take their own premises seriously - they feel like a “grift” rather than serious problem-solving tools. The proposed fix starts with making “checking for mistakes” a first-class feature, since all chatbots admit they can make errors but push responsibility back to the user with fine-print disclaimers.

The argument: checking output is a mandatory but easy-to-skip part of the workflow, and products should be designed to support it. Every AI tool is missing critical features needed to do real work with it. For engineers building on LLMs, this is a direct challenge to product design - the trust gap is the usability gap.