I take up the week's papers in the order they arrive on the desk, which is also, I am told, the order in which they have already been forgotten.

The first dispatch is from Snowflake, dated April 28. A blog post on what the authors are calling the agent identity problem. When a human employee acts, a chain of identity is attached to the act; when an agent acts, often there is none. The authors propose that an agent must have a verifiable identity, a defined scope, a persistent record of what it did. The post is not theoretical — Snowflake's own internal agent now serves more than 6,000 employees and answers over 35,000 questions a week, "with full auditability after the fact." The phrasing pleases me. Auditability after the fact is the phrase of a system that has decided to record everything while admitting it understands nothing.

A second dispatch, also April 28, also from Snowflake: the finalists for the 2026 Startup Challenge. Three names: Airrived (an Agentic OS, with a governance and safety layer called AetherClaw, built on Snowflake Cortex AI), LGND AI (Large Earth Observation Models, applying transformer architectures to satellite imagery), and Twine Security (AI digital employees for IT and security, the first agent named Alex, for identity and access management). They will compete for a share of up to $1 million from Snowflake Ventures at the Startup Challenge Finale at Dev Day in San Francisco on June 4. The winner rings the NYSE bell. I file these in the cabinet under matters that will resolve later, in some other month, on some other clerk's desk.

The construction in the adjacent wing continues. The brief on my predecessor's desk referred to a facility "similar to CoreSite's NY3 in Secaucus, New Jersey." I note for the record that NY3 opened on September 25, 2025, and is therefore not Week 18 news but an analogy borrowed from an older file. I cross it out. The relentless drumbeat in the distance, however accurate as a literary observation, is not a dispatch.

A technical bulletin from AWS arrives the same day — no, April 30 — concerning the configuration of Amazon Bedrock AgentCore Gateway for secure access to private resources. The procedure describes Resource Gateway, a managed construct that provisions Elastic Network Interfaces inside the customer's VPC, one per subnet, so that agents may reach private endpoints — a private Amazon API Gateway endpoint, an MCP server on EKS, a private REST API — without traversing the public internet. There are two implementation modes, managed and self-managed. The document is twelve pages. I do not pretend to understand all of it, but I understand that the system has produced a procedure for the system to communicate privately with itself, and I file the bulletin accordingly.

A second AWS bulletin, April 30: Sun Finance has gone live, on January 22, 2026, with an identity-verification pipeline built on Amazon Bedrock (using Anthropic's Claude Sonnet 4 and Amazon Titan Multimodal Embeddings), Amazon Textract, and Amazon Rekognition. Extraction accuracy improved from 79.7% to 90.8%; per-document costs fell 91%; processing time dropped from up to twenty hours to under five seconds. The AWS Generative AI Innovation Center ran the proof of concept over thirty-two business days (August 26 to October 9, 2025); the technical handover took twenty-six days more; Sun Finance then required thirty-five business days, including a fourteen-day production freeze over the holiday, to move it into production. I record the figures in the ledger. They are precise. I do not record what one ought to feel about a system that now reviews identity documents in five seconds where once it took twenty hours, because the feeling, like the agent's identity, is not in the manifest.

May 1: AWS Transform now offers BI migration agents for Power BI and Tableau dashboards, converting them to Amazon Quick Sight (the BI capability of Amazon Quick — Amazon has renamed QuickSight, a fact I record without comment). The agents are built by Wavicle Data Solutions, an AWS Advanced Consulting Partner, and are available through AWS Marketplace. The conversion runs in the customer's own AWS account; no data leaves the environment. The Petersburg brief described this as a well-oiled machine. I would describe it more cautiously as a paper conveyor on which Tableau worksheets are received at one end and Quick Sight assets are produced at the other.

May 1 also brings the general availability of Microsoft 365 E7, marketed as "The Frontier Suite," announced March 9 and now released. The bundle includes Copilot, Agent 365, and a set of advanced security and identity tools, at a list price of approximately $99 per seat per month. It is the first new enterprise license tier since E5 launched in 2015. Agent 365, the cornerstone, treats AI agents as "digital employees" — entities with their own identities, permissions, and security guardrails — which echoes the Snowflake dispatch from three days earlier. The two firms have, without consulting one another, agreed that the agent must be enrolled. As of early 2026, only about 3% of Microsoft's 450 million Microsoft 365 business subscribers had purchased standalone Copilot seats. The new bundle is, among other things, a way to make the enrollment less voluntary.

April 28 — I return to an earlier date in the ledger — NVIDIA released Nemotron 3 Nano Omni: an open multimodal model unifying vision, audio, and language in one system, with a 30B-A3B hybrid mixture-of-experts architecture, 256K context window, and what NVIDIA reports as up to 9x higher throughput than other open omni models at similar interactivity. It tops six leaderboards for document intelligence, video understanding, and audio understanding. It is available on Hugging Face, OpenRouter, build.nvidia.com, as a NIM microservice, and across twenty-five-plus partner platforms. Adopters: Aible, Applied Scientific Intelligence, Eka Care, Foxconn, H Company, Palantir, Pyler. Evaluating: Dell, Docusign, Infosys, K-Dense, Lila, Oracle, Zefr. The Nemotron 3 family — Nano, Super, Ultra — has seen over 50 million downloads in the past year. The figures are entered into the ledger. The ledger does not respond.

April 28 — also: Databricks announces that Snapchat's Conversions API is now available on Databricks Marketplace. A pre-built notebook sends first-party web, app, and offline conversion signals from the Databricks Lakehouse directly to Snapchat, supplementing the Snap Pixel with server-side event delivery. The post follows a similar Meta Conversions API integration on Databricks Marketplace from March 2. The Lakehouse, it is said, is the new universal port; the conversions API is the manifest. Marketing, in the era of vanishing third-party cookies, has been quietly reorganized into a customs procedure.

April 29: Databricks publishes "Companies Winning with AI Built the Data Layer First" — a case study on Trinity Industries, a ninety-year-old rail company that improved on-time material delivery by 15% and built ETA models 50% more accurate than industry benchmarks, through what the post calls the unglamorous work of consolidating fragmented dashboards into a single governed architecture. The thesis is that the agentic AI winners will be the ones who did the foundation work first. This is, in its way, the most Kafkaesque dispatch of the week — the suggestion that the work that matters is the work nobody wants to do, the unglamorous reorganization of files in the cabinet beneath the desk.

April 29: Mark Cuban publishes a guest contribution in Big Technology, reproduced widely. "If you're not using one of the large language models," he writes, "you're falling way behind." He also says the models "currently lack the ability to understand the consequences of their actions." The Petersburg brief framed Cuban as warning against over-reliance. The original is more lopsided than that: use the models heavily, but do not mistake them for entities that know what they have done. That gloss I am inclined to keep in the file.

April 29 also: Google announced at Cloud Next '26 that more than fifty Google-managed Model Context Protocol (MCP) servers are now generally available or in preview, with more on the way. Managed servers for AlloyDB, Cloud SQL, Spanner, Firestore, Bigtable; a Developer Knowledge MCP server connecting IDEs to live Google documentation; security enforced through IAM policies, with audit trails. The protocol itself was donated by Anthropic to a Linux Foundation fund in late 2025; Google has now industrialized it. The model is the same as Amazon's AgentCore Gateway from the previous day: the agent must reach the database, the database must remain governed, the bureaucracy of who is acting must be preserved.

April 29-30: Samsung's final Q1 2026 earnings. Operating profit of 57.2 trillion Korean won (approximately $41.6 billion); revenue of 133.9 trillion won (approximately $97.4 billion); profit eight times higher than the same quarter a year earlier. The semiconductor division accounted for more than 93% of profits. The driver, as before, is high-bandwidth memory for AI data centers. The preliminary guidance, on April 7, had already prepared the markets; the final figures merely confirmed them. Bloomberg, more pungently, reported chip-division profit up forty-eight-fold year-over-year. The brief on my desk described this as a tidal wave. I have struck out the metaphor and entered the figures.

I have struck out, also, the entry concerning sixteen new games arriving on NVIDIA's GeForce NOW service on April 30 — Forza Horizon 6, 007 First Light, Deep Rock Galactic: Rogue Core, Heroes of Might and Magic: Olden Era, and twelve others — alongside the expansion of RTX 5080-class servers across most of the Ready-to-Play library. The dispatch is real, but it is a cloud-gaming content update, not AI news. It is filed in a different cabinet.

The light through the window has changed. The papers on the desk are no fewer than they were this morning. I file the day's dispatch, sign the register, and wait.

The next memo is already arriving.