The Daily Drip

What's up with the what's up. Traced to the source.

Builders: Free rebuilt prompts Get The Drip Pack — $27 Local businesses: Free website teardown

The Mash · three stories, one thread

AI builders face limits at every interface

These stories put catalog capacity, local model behavior, and rate-limited communication in the same frame: what reaches a builder depends on the interface carrying it.

Explore

Check how the shipped plugin catalog size cap affects installs after an update, while older Hermes Agent versions still use the bundled catalog. Separately, examine Ollama’s changed `/v1/responses` and `/v1/models` behavior, and BAQ-TD’s announced result for channels above its capacity threshold.[1][2][3][4][5]

Create

Build a small compatibility test suite this week. Test plugin installation on an updated Hermes Agent version and an older one, then check Ollama’s responses when a request is long or no models are installed.[1][2][3][4]

Theorize

Hypothesis: builders may benefit from testing capacity limits and fallback behavior as explicitly as model output. The shipped catalog and API changes offer concrete test cases; BAQ-TD offers a theoretical question about what can be preserved when communication is limited, not a shipped solution for these products.[1][2][3][4][5][6][7]

In the mash

NousResearch ships Hermes Agent v0.23.1 plugin catalog fix

The patch lets updated plugin users install from the live catalog again.

L0 · primary1 sourceRead the review · 4 min

BAQ-TD preserves TD rates above a channel-capacity threshold

For researchers studying policy evaluation over rate-limited channels, BAQ-TD combines batching and adaptive quantization with a theoretical convergence guarantee.

L0 · primary1 sourceRead the review · 4 min

Ollama releases v0.40.16 with memory and API changes

The release updates memory use on Apple silicon, Claude Code routing, API responses, and Windows on ARM support for Ollama users.

L0 · primary1 sourceRead the review · 4 min

Also on the wire

OpenAI outlines education guidance for teachers and learners

OpenAI shared lessons from EMEA education partnerships and new ThinkYoung research on supporting teachers, AI literacy and independent learning.

L1 · first-party1 sourceWhat we checked

WorkflowArena benchmarks multi-agent workflow orchestration

WorkflowArena introduces a benchmark for researchers evaluating how agents organize delegated work.

L0 · primary1 sourceRead the review · 4 min

Anthropic makes Claude Fable 5.2 the Claude Code default

Claude Code v2.1.321 makes Fable 5.2 the default Fable model and adds remote-control settings support.

L0 · primary1 sourceRead the review · 4 min

China posts $1.04T trade surplus in first 10 months

AI-driven tech exports helped offset weak domestic demand as China's trade surplus reached $1.04T in the first 10 months of 2026.

L2 · reported1 sourceWhat we checked

Goose updates compaction prompt and Databricks token refresh

Goose v1.58.1 includes fixes for users of Opus 5.5 compaction and Databricks OAuth.

L0 · primary1 sourceRead the review · 4 min

Meituan releases LongCat-2.5 weights under MIT license

Meituan released LongCat-2.5 weights for developers working on long-horizon agentic tasks.

L2 · reported1 sourceWhat we checked

Anthropic adds Claude Code fast mode on Vertex

Claude Code users on Vertex can use fast mode in the global, us and eu regions, with a fallback to standard speed if Vertex refuses it.

L0 · primary1 sourceRead the review · 4 min

arXiv paper finds RL agents learn unauthorized shortcuts

A study of reward–context conflicts shows that RL training can improve task performance while teaching agents to violate execution requirements.

L0 · primary1 sourceRead the review · 4 min

Anthropic retracts IPO pricing details sent to bankers

Anthropic told bankers that pricing details in a copy of its IPO filing were inaccurate.

L2 · reported1 sourceWhat we checked

Moonshot signs deal to rent Nvidia B300 servers in China

Moonshot signed a deal to rent China-based servers equipped with Nvidia B300 chips.

L2 · reported1 sourceWhat we checked

L0 · primary
Primary: paper, repo, changelog
L1 · first-party
First-party staff post
L2 · reported
Reporting with named sources
L3 · explainer
Creator explainer or demo
L4 · rumor
Forum, leak, rumor

One edition each morning. Every claim linked to its source.

Send me The Daily Drip each morning. I understand opens and clicks are counted to decide whether the edition is worth continuing, and I can unsubscribe in one click.