October 14, 2026 · 5:12 PM PT edition
Ramp's 50x question flood, Opus writes the plan, and a 278,000-star Claude Code setup
Fivetran and dbt Labs on why MCP is not a data architecture, a cheaper way to run agent swarms, and DeepMind hiring for scaling laws.
When Ramp opened an AI research agent to the whole company, it got 20 to 30 times the usual number of data questions within weeks. Meanwhile a new Claude Code trick has Opus 5.5 write the plan as a script and hand it to a swarm of cheaper agents.
Here is what happened in AI and DevOps this cycle.
George Fraser and Tristan Handy, who now lead Fivetran and dbt Labs together, sat down with a16z’s Martin Casado to talk about open data infrastructure in the age of AI agents. Fraser opened by naming Stripe, Ramp and DoorDash as companies that built their own AI analysts on top of their data stacks. The wildest number on the panel: Ramp’s Slack research agent drew 20 to 30 times the questions in a couple of weeks and 50 times within six months. The panel calls it latent demand, and Fivetran has its own proof. It wired Claude to an MCP server sitting on its dbt-modeled data, and in one meeting a colleague’s guess of 8% and 4% turned out to be 16% and less than 1%. The vibes in the room, the panel says, are often very wrong.
The same panel took a hard line on a popular shortcut. Casado argues that MCP is an access protocol, a bit like HTTP, with nothing to say about structured data, guarantees, security or formats. So pointing an agent at the Salesforce, Workday and Jira MCPs is not a data architecture, and the Fivetran side adds that it is a permissions mess and incredibly token hungry and expensive. Their recipe is to keep your own copy of the data in an open table format, with Iceberg as the one everyone supports. Handy’s point for anyone feeling impostor syndrome: agentic analytics is about 80% the work you’ve already been doing.
There is a fight brewing over who owns the data agents read. The Fivetran and dbt side says some SaaS vendors are talking about tollgates on AI access to customer data. It points to Amazon charging sellers for the Selling Partner API, the feed of what they sold, a fee that lasted only a couple of months before the outrage killed it. Casado reads it as the old Win32 and web API tension again, a fork where vendors either follow the new consumption layer or fade. He also rejects the doom version, saying that across hundreds of companies he sees this used for productivity, not headcount reduction.
The plan-as-script trick comes from Anthropic, according to a write-up from delost. Opus 5.5 writes the plan once as a script, and a swarm of cheaper agents carries it out. Because the script holds the plan, Opus burns far fewer tokens while the job gets bigger. If it holds up, frontier-model cost stops scaling with the size of the task.
Claude Code add-ons are getting enormous. Suraj Sharma says one developer’s personal setup, Everything Claude Code, has passed 278,000 stars and 41,000 forks, is MIT-licensed and ships 71 agents for planning, coding and more. Yarchi points to OpenMontage, which turns Claude Code into a video studio with 12 production pipelines, 100+ tools and 700+ skill files, and pitches it as a free replacement for the AI influencer stacks agencies charge thousands for. Those counts come from the promoters, so check the repos before you install.
Corey Ganim shared a copy-paste prompt that connects ChatGPT and Claude subscriptions to Grok Bot so every bot can draw on them. Step one is installing the Claude Code CLI and the OpenAI Codex CLI, upgrading Node if needed. His “Elon approved” tag is a promoter’s flourish, but the idea is real: one assistant brokering the model plans you already pay for.
Cody Schneider argues that if a human can do it on a computer, an agent can do it, and says marketing is moving fastest, with research, campaign launches, media buying and growth loops all handed to agents. A clip making the rounds quotes Andrej Karpathy saying “99% of my coding routine is automated by AI agents” and calling agentic harness engineering the level past vibe coding. That quote comes secondhand and wasn’t checked against the original 30-minute talk.
Spencer Frei is hiring a Student Researcher to work on scaling laws at Google DeepMind. It is a 19-week, in-person project in downtown San Francisco, ideally starting in January 2027. A small listing, but it shows the people mapping how models grow are still adding hands.
That’s the cycle. Go put your data in Iceberg, let Opus write the plan, and see you tomorrow.
Also on the radar: these also appeared this cycle, but there was no transcript to report from yet (captions rate-limited by YouTube, no English captions, or no speech), so they aren’t summarized here: I Spent Two Years Building the Median — Hariprasad Sudharshan (AI Engineer); Exclusive INSIDE the Secretive AWS Networking Lab (ServeTheHome); Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder (AI Engineer); How to Actually Tokenmaxx and Save Millions — Alvin Pane, OutRival (AI Engineer); How stacked PRs work behind the scenes at GitHub – S02E05 – The GitHub Podcast (GitHub); Rebuilding CI for AI-Scale Commit Volume — Gautam Korlam, SonarSource (AI Engineer); The Cost of Accessibility in AI — Jemima Abu, CAIS (AI Engineer); What Happens When AI Gets a Human Body? (Core Memory Podcast); Your Fridge Is a Valid Payment Device — Eleonora Drykina & Nina Žižakić, In The Pocket (AI Engineer); Gemini 4 Argon: 10 incredible AI demos (1M token output in action) (Google Cloud Tech); How to onboard to Looker using your favorite LLM and skills (Google Cloud Tech); The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale (AI Engineer); Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers (AI Engineer); Gemini 4 Argon Puts Google Back On the Frontier (Every).