The Collection

Marcus’s Sunday letter is short on purpose. Annual Recurring Revenue is the thing you want: subscriptions, next year likely the same or better. It recurs. Annualized Run Rate can mean as little as the best month times twelve, then a hope that this month becomes next year’s average. That could happen. Or the best month is not typical, and you never print a year that matches the cover.

When people tell you Anthropic had this or that ARR, he writes, you need to know which meaning they are talking about. They rarely tell you. They really mean the latter. He names the pressure: the kind of revenue they were making in the second quarter of this year may well not recur, at the same scale, now that companies like ATT are turning to open-source models from other vendors. He dates that turn to the aftermath of tokenmaxxing’s collapse. He does not invent ATT’s bill. He does not invent Anthropic’s.

Annual Recurring Revenue recurs. Annualized Run Rate does not have to.

Then 1995. Netscape had revenue, doubling every quarter, a public company worth billions without a profit. Later that year Microsoft offered a competitor for free. Netscape was never the same, all but out of the game a decade later. Caveat emptor, he closes. The Collection’s job is smaller: when someone boasts about ARR, print the question. Which one.

Tom Krazit’s Saturday Runtime opened on a Product Saturday: Cursor unveiled a version-control host in the same week GitHub kept struggling to stay online. Origin, in Cursor’s own docs as he quotes them, is a git forge for storing and sharing code. Host repositories. Sync projects from GitHub. Browse the team’s Origin repos in the browser. Git is the second-most famous open-source project Linus Torvalds ever developed. GitHub has been the most widely used host around it since 2008.

2008 was a long time ago. Anyone shipping in 2026, Krazit writes, knows GitHub is straining under AI-driven software from agents. Vicent Marti, in the launch post, walked the early GitHub architecture. Most of those decisions made sense then. They are coming back under a load you cannot throw compute at. Continuity is Cursor’s Git storage: learn what Spokes did well, fix what years of production now show as problems. Agents made the situation worse. Origin is built around AWS S3 and write-ahead logging.

The 2008 git assumed people. The 2026 git has to assume agents.

GitHub’s own receipt arrived in the same pile. Krazit flags a 17 August outage post from GitHub CTO Vlad Federov, thoughtful and contrite, promising to do better, and a note from Airbnb’s Lorin Hochstein on autoscaling and the component-substitution fallacy in the cascade. We have those titles. We do not reprint the post-mortems as if we sat the incident. The Collection prints the fork: a forge built because the old one is breaking, in public, on a Monday.

NVIDIA’s general-purpose coding agent AVO scored a perfect 100.00 RHAE on the ARC-AGI-3 public set. It cleared all 183 levels across all 25 environments with no instructions, rules, or stated goals. The weekly is careful with the next line. Claude Opus 5 alone scores roughly 30 percent on the same benchmark. Wrap it in AVO and it goes to 100 percent. NVIDIA frames that as system design, not model capability alone, unlocking long-horizon work.

The loop is named. Inspect context, plan the next change, implement, evaluate with a scoring function, diagnose and repair from failed attempts, commit accepted candidates into a growing solution lineage. Persistent memory keeps prior implementations, evaluation results, compiler output, and reasoning, so the agent resumes from current state. A separate monitor watches for stagnation and redirects. The same architecture was built for CUDA GPU kernel optimization and moved to interactive reasoning unchanged, using about 12 percent fewer environment actions than the prior VISTA baseline.

The model at 30 percent. The harness at 100. Do not confuse them.

TrueFoundry, the same issue, open-sourced TrueForge under an MIT license: a vendor-neutral harness, MCP tools, a skills registry, sandboxed execution, approval gates, subagents, durable state, step-level tracing. It runs on your iron against any model. The TrueFoundry AI Gateway is optional. On DevRev’s Enterprise-Bench, evaluated blind, they report roughly 30 percent lower cost on identical tasks with the same model, and up to 75 percent savings when routing to open-source models with accuracy matched. The repo has cleared 2,800 stars since the 19 August launch. We have their numbers. We did not sit the bench.

Ken Huang’s Saturday letter is a crosswalk, not a rant. He maps the Model Context Protocol against the six domains of Proof-of-Control. MCP is hosted by the Linux Foundation, JSON-RPC 2.0, the place an agent reads data it did not author, calls a tool it does not own, and takes an action outside its own process. Every such crossing is a place evidence goes missing. His finding is not that the protocol is sloppy. It is that the controls are thorough and the record is held by the party being checked.

Revision 2026-07-28 is stateless. No initialize handshake, no session. Every request carries version, capabilities, client identity, authorization in its _meta field. That is good for a per-action evidence record. It is still Tier 2: attestation. Token validation, scope enforcement, consent, schema: all real, all normative, all attested by the operator who runs the server. Moving a spec-conformant deployment from Tier 2 to Tier 3, he writes, requires no new controls. It requires changing who produces and holds the record.

A careful log written by the operator is still a month. It is not a year you can audit.

He works an example. A support agent running Opus 5 issues a $42.50 refund on a billing MCP server. Six weeks later an insurer’s adjuster asks whether the agent was authorized and stayed inside the limit. The Tier 2 row is not weak: audience-bound tokens, PKCE, scope, schema, per-client consent. The adjuster still has nothing they can rely on without trusting the company that wrote the log. That is the binary threshold. We print the threshold. We do not print a refund that did not happen.

Salesforce shipped Slack Code in the same Runtime. Anyone with access, not just developers, can @-mention coding agents like Claude Code or Codex inside Slack. Slack Code opens a new channel for that project. The team can see it and comment as the work unfolds. The company’s line, as Krazit quotes it: a dedicated place to work with an agent on a specific coding session or project, together, right in Slack.

That is the forge’s other half. Origin stores the repo under a load git was not built for. Slack Code makes the agent session a room other people can walk into. Neither one is a year of recurring revenue. Both are infrastructure for a month that is suddenly very busy. If the run is a projection, the channel is where you watch whether it actually ships.

The agent used to work in a private terminal. Now it has a channel.

Krazit also notes Snowflake adding dynamic model routing in Cortex AI Gateway, and Cerebras announcing CS-4: a claimed 30 times performance, up to 10 times more throughput per watt than CS-3. Stripe and OpenRouter we already printed on Wednesday. The hangar we printed yesterday. Those numbers stay on those shelves. Today’s paper is the trick, the forge, the harness, and the log.