8 min read · 11 sources

Founders have always worked long hours. What’s new is that the thing they’re building is outpacing them. AI agents are moving so fast — capabilities expanding, model updates landing weekly — that founders feel compelled to keep up or lose precious hours of work. The result is erratic sleep, an inability to focus on anything else, and a pervasive sense that the pace is unsustainable but stopping isn’t an option. This is the defining work-culture story of 2026, and it has consequences beyond personal burnout: startups running on exhausted founders make worse decisions, miss compliance obligations, and ship half-tested features.

The DOJ probe may force VCs to give up board seats or decline them entirely, changing term sheets for anyone raising now.

DOJ Opens Antitrust Probe of Andreessen Horowitz

Source: techcrunch.com ↗

The Department of Justice has spent close to a year investigating Andreessen Horowitz over board seats the firm holds in competing AI companies, per a Bloomberg report cited by TechCrunch. Ben Horowitz sits on the Databricks board; Martin Casado holds a seat at Fivetran. Regulators are asking whether a VC firm that holds seats on both sides of a competitive market has crossed an antitrust line.

The VCs TechCrunch interviewed seemed genuinely baffled. One question is why this rises to the DOJ’s priority list given a16z’s documented ties to the Trump administration and the firm’s relative silence on policy during the Biden years. Nobody outside the probe knows what the underlying allegations are. What engineers at VC-backed startups should watch: if firms respond by giving up seats or declining them, board composition, information rights, and term sheets change for anyone raising now. The probe may set an “example” that pressures smaller VC firms to audit their own board conflicts — particularly in AI portfolios where portfolio companies increasingly compete directly.

Building a Semantic Layer at PostHog

Source: x.com ↗

PostHog published a detailed walkthrough of a problem every data-savvy company eventually hits: ask Claude, Cursor, and PostHog AI the same metric question — “what was our MRR last month?” — and get three different answers. The root cause isn’t the models. It’s that the definition of MRR, which table is authoritative, and what joins are required all live in tribal knowledge rather than a single reference.

The fix is a semantic layer — a dictionary-style layer that sits on top of the existing data warehouse without copying data or replacing anything. It encodes metric definitions, trusted tables, and join semantics so agents and humans read from the same source of truth. PostHog’s post walks through how they built theirs.

For engineers, this matters because agentic analytics are only as reliable as the shared definitions agents read from. Without a semantic layer, even a better model keeps producing plausible-but-wrong metric answers. If you’re shipping AI-assisted data tooling, this is the unglamorous plumbing that determines whether outputs are trustworthy.

Build vs Buy When Building Just Got Cheap

Source: kevingoldsmith.substack.com ↗

Kevin Goldsmith recounts approving an open-source auth stack at an early-stage B2B startup to save on vendor costs. A one-week extension became constant security patching, building features a vendor would have shipped, and lengthy security-review conversations with enterprise prospects. Roughly 30 days after the company moved to a vendored solution, a serious vulnerability in the open-source component compromised several sites that hadn’t migrated.

Goldsmith frames this around a team mantra — “let’s not invent things that already exist” — with deliberate absolutism to set a high bar. The lesson isn’t that open source is bad; it’s that build-vs-buy math that ignores ongoing maintenance, security patching, and sales-enablement costs routinely comes out as a wash at best. Luck is not a strategy.

X Ads MCP: AI Agents Running Ad Campaigns

Source: threadreaderapp.com ↗

X launched an Ads MCP that lets Grok, Grok Build, Claude Code, and other AI agents connect directly to X Ads accounts and manage campaigns through normal conversation. Agents can pull real campaign data, analyze results, create and manage campaigns, search interests and locations for targeting, add or remove targeting, create ad posts, promote existing posts, check performance, and update and activate campaigns. There are 23 X Ads tools available through the MCP.

This is a concrete example of the MCP (Model Context Protocol) pattern going mainstream: instead of logging into a dashboard, you tell an agent what you want and it handles the API calls. For engineers building agentic workflows, the MCP surface area is growing fast, and understanding how to expose your own product via MCP is increasingly relevant.

Notion Ships a TypeScript Workspace API — and Finds Most Teams Still Using AI Superficially

Source: threadreaderapp.com ↗

Notion announced “Notion as code” in beta: defining a whole workspace — teamspaces, databases, custom agents — in TypeScript and deploying via API for git-versioned, reproducible setups. The company also shipped a new HTML block for embedding interactive prototypes, dashboards, and explainers directly in pages.

In the same announcement, Notion published survey data from 6,000 professionals across 10 markets: only 2% of organizations are operating at the most advanced level of AI adoption, with 88% still using it superficially. The report frames AI maturity as levels that deepen rather than graduate from.

For engineers, the TypeScript-defined workspace API changes how Notion setups can be coded, version-controlled, and shared. If you’re managing a multi-team Notion deployment, this is a real infrastructure upgrade. The survey data is useful context when making the case for deeper AI investment internally.

Why Your Local LLM Feels Dumber Than It Is

Source: forum.level1techs.com ↗

The Level1Techs forum post arguing local LLMs feel “dumber” than reference benchmarks is worth reading for the data. Someone ran 44.8 terabytes of tests on the same model, same question: right answer on one GPU, wrong answer split across two, right again across four. One common compression setting quietly broke the model’s ability to call outside tools; a slightly bigger one worked fine.

The root issue is implementation-specific inference hazards. Different GPU generations, instruction sets, and software stacks compute logits and token probabilities differently. Standard prompt-and-pray evaluation is a poor proxy for agentic workloads. The author recommends running representative benchmarks — terminal-bench, HLE, SWE-bench-style, MMLU, long-context tool-calling, domain-specific evals — rather than zero-shot test prompts, and warns against judging heavily quantized GGUF runs as representative of model quality.

Cloudflare Bot Preference Sync Hits All Plans

Source: blog.cloudflare.com ↗

Cloudflare launched Bot Preference Sync, available to all customers from Free through Enterprise. It mirrors the AI bot configuration set in Cloudflare’s dashboard into the site’s robots.txt automatically, and can be toggled on or off at any time. The feature addresses a common mismatch: robots.txt disallows a crawler but enforcement rules don’t block it — a gap some crawlers treat as license to ignore stated preferences.

It builds on Cloudflare’s July 1, 2026 launch that introduced separate preference controls for Search, Agent, and Training traffic, plus earlier managed robots.txt entries and edge-enforced blocks against Training crawlers. For engineers, keeping robots.txt and WAF/bot-management rules in sync is now automatable across every plan tier, and AI Search/Agent traffic is being treated as a first-class category distinct from Training crawlers.

Is Agentic: A Public Site Scanner for AI Agents

Source: is-agentic.com ↗

Is Agentic is a public site scanner that scores how “agentic” a website is based on what AI agents can discover, retrieve, understand, and use. The score splits into Essential checks (server-rendered content, correct HTTP behavior, clear document structure, recoverable errors, usable controls) that carry most of the weight, and Recommended checks (API, OAuth, GraphQL, MCP server, developer portal, commerce surfaces) that activate only when applicable. Reports include an observed agent journey, are available through a public JSON API and an MCP server, and are rendered server-side at stable URLs so crawlers don’t need JavaScript.

For engineers, the tool gives a concrete, machine-readable checklist for making sites agent-friendly and surfaces remediation evidence per finding. If you’re responsible for site discoverability or building agentic workflows that depend on site scraping, this is a useful benchmark.

You Have Never Been Your Own Customer

Source: sixteenventures.com ↗

A sharp post from Sixteen Ventures argues companies consistently fail to find their own customer-experience problems because no one inside the company has actually been the end-to-end customer of the systems running in their name. Two examples: a fear-based disclosure form wrapped around an insurance upsell that frightened a realtor’s client into a slower process, and an eye-care clinic whose online booking was overturned by an email requiring a phone call within a window that had already passed, producing a no-show charged to the patient.

The takeaway: vendor-skinned software with your logo on it is functionally you, and the step of indirection is exactly what makes the harm invisible. For engineers and product owners, the point lands hard: shipping a feature isn’t the same as having personally exercised it as a customer. Unexercised systems keep producing self-inflicted churn the company never sees.

The GTM Playbook Has Shifted More in Two Years Than the Prior Decade

Source: growthunhinged.com ↗

Kyle Poyar’s “Growth Unhinged” field guide catalogs six core GTM motions — inbound, outbound, product-led, account-based, paid digital, partners — with inbound currently the most popular per a State of B2B GTM report. The vocabulary has shifted: AEO (Answer Engine Optimization), context engineering, MCP, FDE, and token economics are now load-bearing terms. Inbound is most adopted at 23%, product-led still dominates below $1M ARR and under $5K ACV, and 47% of teams are testing AI discovery channels. Hybrid pricing sits at 37%, and AI credit models are forecast to grow 114% in twelve months. On hiring, 39% of top AI companies are filling forward-deployed roles at $213K to $370K+.

For engineers building GTM tooling or pricing infrastructure, the vocabulary shifts correspond to concrete product and pricing changes — particularly around agentic surfaces (MCP) and AI-aware discoverability (AEO). If you’re shipping anything in the go-to-market tooling space, this field guide is a useful landscape map.