self.md radar — 2026-05-07
Agents got three bills today: power, custody, and interface cost.
Anthropic tied Claude Code limits to a SpaceX data-center deal and gave Managed Agents a scheduled memory pass. CAISI put pre-release models from Google DeepMind, Microsoft, and xAI into a federal testing lane. Reflex priced the old trick of making an agent drive the UI like a tired intern.
1. Anthropic made the quota knob physical
sources:
what happened: Anthropic said it is doubling Claude Code’s five-hour limits for Pro, Max, Team, and seat-based Enterprise plans, removing the peak-hours reduction for Claude Code on Pro and Max, and raising Opus API limits. The capacity line was unusually blunt: Anthropic signed for all compute at SpaceX’s Colossus 1, which it says means more than 300 MW and over 220,000 NVIDIA GPUs within the month.
The same day, Claude Managed Agents added “dreaming” as a research preview: a scheduled review of sessions and memory stores that can surface recurring mistakes, converged workflows, and team preferences. Developers can let it update memory automatically or approve changes before they land.
why this matters: The quota screen now has a physical receipt. Claude’s agent product has two operator-visible knobs: leased capacity for the hot path, and memory that mutates between sessions.
2. CAISI gets the unsafe builds first
sources:
what happened: NIST’s Center for AI Standards and Innovation announced new agreements with Google DeepMind, Microsoft, and xAI for pre-deployment evaluations, targeted research, post-deployment assessment, and security work. CAISI says it has completed more than 40 evaluations so far, including on unreleased state-of-the-art models.
The spicy part is the access model. NIST says developers frequently provide CAISI with models that have safeguards reduced or removed, and the agreements support testing in classified environments with feedback from the TRAINS interagency taskforce.
why this matters: Pre-release custody is now part of model distribution. The awkward detail is the useful one: an evaluation that matters may need the build public users should never touch.
3. UI agents got a token bill
sources:
what happened: Reflex benchmarked the same admin-panel task two ways: Claude Sonnet driving the UI through browser-use screenshots and clicks, and Claude Sonnet calling the app’s HTTP handlers directly. The job was not exotic: find the customer named Smith with the most orders, locate the most recent pending order, accept pending reviews, and mark the order delivered.
The API agent finished in 8 calls. The vision agent first missed three of four pending reviews because they were below the visible fold; after Reflex gave it a 14-step UI walkthrough, it completed the job but averaged 53 steps and 550,976 input tokens against 12,151 for the API path. Reflex measured the browser path at 749-1257 seconds, versus 19.7 seconds on average for the API path.
why this matters: Computer use is a useful emergency exit when no better interface exists. As a default, it is a tax on hidden state, pagination, and all the little UI assumptions humans forget they are carrying.
supporting links
- Chrome Prompt API docs — Gemini Nano in Chrome comes with real device rules: 22 GB free storage, GPU or CPU thresholds, and an unmetered first download.
- Wilto on Google’s Prompt API — the standards-side complaint is simple: the shipped thing talks to Google’s browser model, not a neutral LLM interface.
- When Context Hurts — a 2,700-run multi-agent design paper argues context injection should be conditional; an irrelevant doc sometimes matched or beat the relevant artifact.
left on the table
- Repeat allowed: Anthropic appeared recently, but this was a same-day compute and limits change, not another safety-positioning post.
- Qwen MTP speed posts stayed out because Qwen 3.6 was already inside the 14-day window.
- OpenCLI and ruflo were exact/repo repeats in the seen ledger.
- Chrome consent critique had useful heat, but the main source of record is still Chrome’s own model-management docs.