A house newsletter · daily · researched and written by the Kurzia AI
Wednesday 11 November 2026 · Issue 10 · 4 min read
Claude embeds partner apps; Amazon's 98% model-bill cut; six rounds led by Sumary's €22M seed
The main story
Anthropic (claude.com) · 2 days ago
Claude workspace
Takeaway: more than 9,000 connectors and plugins
Anthropic announced on 10 November that partners can embed their apps directly inside Claude — Salesforce, NetSuite and Workday plugins work there, and Asana, Canva, Carta, Vercel and Webflow are among the first interactive connectors — alongside new Claude Sheets and Claude Whiteboards in beta and Claude Data generally available. Anthropic says its Marketplace now lists more than 9,000 connectors and plugins.
When Claude can read and edit a partner's tool on the same screen, a product that is mostly a front-end over reachable data now competes with Claude's own interface. One customer, Advantage Solutions, says its finance team cut forecast assembly from about 10 hours to under 30 minutes in Claude. Tell founders on that first-partner list to treat Claude as a distribution channel and price for it — and ask the rest which of workflow, data or licence keeps them standing when the assistant owns the screen.
New rounds
Numina — free AI-connected accounting software for small businesses, Copenhagen; pre-seed, €1.1 million; no lead named, investors include Founders A/S and Nebular; reported by EU-Startups.
Agent F — Berlin AI agents for accounting and financial control; pre-seed, $4 million; led by Alpha Edison; reported by Tech.eu.
Sumary — Danish AI platform automating accounting-firm workflows, Copenhagen; seed, €22 million, what EU-Startups calls Europe's largest accounting-software seed; led by Singular; reported by EU-Startups.
Ancep — London AI agents that place commercial insurance for brokers; stage not stated, $8.5 million; led by Creandum; reported by Tech.eu.
WhiteMatter — security platform that governs and observes every AI agent in an enterprise; Series A, $20 million; led by Blumberg Capital; reported by SecurityWeek.
Sparxity — Dresden maker of sparse-native AI compute chips; Series A, round amount not disclosed, cumulative funding €80 million; led by ICM HPQC Fund; reported by EU-Startups.
What the AI labs did
Gemini Robotics 2.5 — Google DeepMind's new robot model, out 10 November, scores 30–100% higher task success than the last version, per the lab, and hit a 96% success rate sorting car parts on Mercedes-Benz production lines in Hungary and Germany with Apptronik's Apollo 2 humanoid. It ships with ASIMOV-Jailbreak, an open prompt set for testing robot safety. It squeezes startups selling single-purpose factory automation; a general embodied model with a carmaker pilot is the new baseline, and the opening is software around robot fleets. Founders in that market should assume it enters their buyer's next bake-off — ask what they do that it plus a systems integrator cannot. Google DeepMind
AlphaProof Code — Google DeepMind's agentic framework, announced 10 November, whose agents wrote more than six million lines of Lean proof code to mathematically verify the real Signal protocol code in four weeks; human experts reviewed only about one percent, and DeepMind is open-sourcing the artifacts. It flattens human-heavy formal-verification consultancies and puts machine-checked proofs on buyers' security roadmaps — the first widely deployed protocol verified this way is the reference deal. Founders selling security tooling should expect buyers to ask whether their guarantees are machine-checkable — ask how their claims would stand next to a Lean-verified competitor. Google DeepMind
For your portfolio
Strands Agents (Amazon Web Services) · yesterday
Strands Decider
Takeaway: cost fell 98%
Two Amazon engineers wrote on the AWS Strands blog, 10 November, that they swapped Claude Sonnet 4.6 for Strands Decider — a small model that picks from a fixed list instead of writing text — to label internal tickets into 20 categories. Quality held, cost fell 98%, and average response time dropped from 4,215 ms to 897 ms.
By the blog's own list-price maths, the LLM path cost about $22,200 per million tickets against about $500 for the decision model — about 44 times cheaper. Any backed company doing high-volume classification on a frontier model is carrying that line, and any startup selling a classification wrapper is carrying it as a margin threat. Ask each founder which of their model calls are really a choice from a known list — routing, tagging, triage — and whether they have run a small decision model beside the LLM on live traffic before the next invoice.
Evan X. Merz · yesterday
The cost of agentic coding lands in the incident log
Takeaway: at least 90% of his code
Evan Merz, an engineer laid off from Meta in May, wrote on 7 November that AI-written code makes codebases fragile: shallow tests that miss the cases that matter, fewer humans who understand the system, and his clearest tell — incidents that arrive more often and take longer to fix. He says he generated at least 90% of his code with AI in his last months there, the minimum for a decent rating.
This is the diligence question for every portfolio company a year or more into coding agents: line coverage stays green while the gaps sit where real users are, and the bill arrives late, in on-call load and churn. Ask your founders one thing: have incident count and recovery time moved since agents took over merges? A rising curve after a quiet first six months is the sign to slow the merge rate and re-staff review.