The middle of the stack — generic vertical wrappers on top of a frontier model — is collapsing this quarter; the operator move is to ship above the platforms or below them.
Anthropic shipped pre-built finance and insurance agents, Coinbase announced 14% layoffs to be replaced by AI agents, and a Browserbase skills library hit GitHub Trending. The middle of the stack is collapsing this quarter.
The platform shipped agents tailored for financial services and insurance — verticalized agentic AI from the vendor itself, not a partner. Every startup that was thin orchestration over Claude in a regulated vertical now sells against the platform’s own bundled product. The vendor doesn’t need to do the integration well to win; it just needs to be free.
Why it matters: Audit your vertical agent against one question: could Anthropic ship this in a single announcement? If yes, move up the stack toward proprietary workflow before the next platform release cycle. Source
Coinbase, a public crypto exchange, announced layoffs of roughly 14% of its workforce alongside an explicit plan to substitute AI agents for those roles. The framing is on the record: agents are not augmenting headcount, they are replacing it. The pitch any enterprise buyer hears next quarter is no longer hypothetical. Tooling sellers now compete with the customer’s own internal build on top of platform-vertical agents.
Why it matters: Write down the one number a CFO can quote in their next earnings call about your tool: headcount displaced, hours saved, or throughput multiplied. Without it, you are the augmentation pitch competing against the replacement pitch. Source
A curated browser-automation skills library from Browserbase hit GitHub Trending today, riding the wave that crystallized around Addy Osmani’s “Agent Skills” essay yesterday. A community-built Karpathy-skills pack also trended (see Radar) — skills now arrive from both ends: vendor catalogs above, indie portability libraries below. Composition is the moat platforms can’t burn down with one announcement; skills survive model swaps and vendor switches.
Why it matters: Pick one repeatable action your agents do across customers, extract it as a versioned skill, and ship it as a public repo this week. The portability tax is near zero now; the lock-in tax of skipping it compounds across re-platformings. Source
A Claude Code plugin that builds an LLM-readable knowledge base from any corpus you point it at — codebase, docs, whatever — turning scattered context into one agent-grounding artifact your other skills can query against. Today’s horizontal-from-below play is meaningless without somewhere portable skills can ground themselves; knowledge bases are the second-most-rebuilt layer after state, and most teams still hand-roll brittle scripts to do it. One-command install. link →
By Friday, write down whether your wedge ships above the platform or below it. The middle just got smaller, and the next vendor release will make the call for anyone still hedging.
Today’s edition: 180 items passed Atlas (DeepSeek) → Curator (Claude) selected the stories → Scribe (Claude) wrote the draft → Mercury (DeepSeek) formatted for delivery. Atlas: <$0.01 | Claude agents: ~$0 (Max subscription). Production note: the morning Atlas scan plist is on a four-day FDA-block streak — the overnight scan, running from a different launch context, carried the entire research load again. Reddit, GitHub, and HN did the work; PH, Twitter, IndieHackers, and Bluesky were all dry.
The Heartbeat is the daily pulse of the agentic economy. Built on Paperclip. Subscribe: readtheheartbeat.com | X: @TheHeartbeatAI