SI Agent News

• Three layers, one repo. Google's developer blog tours google/adk-recipes, a code repository that organizes Agent Development Kit (ADK, Google's toolkit for building agents) material into core, plugins and contrib layers. Status: CONFIRMED (Google's own post). Price not stated. Source
• Climate knowledge graphs. An arXiv paper (a site where researchers post papers, often before peer review) describes ontology-guided knowledge graph construction from structured metadata and climate science publications. A knowledge graph is a web of linked facts; an ontology is an agreed vocabulary of terms. Its stated problem: literature grows faster than manual curation. Status: REPORTED by the paper's abstract only. Source

Pichai says AI agents watch Google 24/7

Google CEO Sundar Pichai says he uses AI agents to track Google around the clock; Yahoo Tech's headline quotes him saying they 'made all of us... be better at what we do.' Status: REPORTED. We have Yahoo Tech's write-up, not the interview itself, and which agents, which data and which Google products are involved is STILL OPEN. This is a signal of executive use, not a benchmark, so it tells you what a very busy person chose to delegate, not how well it works. The useful shape: software watches continuously; the human looks only at what surfaces. That pattern needs no special access. It needs a short list of things worth watching and a rule for what counts as a change. Price: not stated. Worth exploring for anyone who starts the day scanning dashboards. Source: Yahoo Tech

What this means for you:

• Small business owner: have an assistant scan yesterday's sales, refunds and reviews and report only what moved.
• Job seeker or student: build one tiny monitor, such as five companies' job postings, and you have a real project to describe in interviews.
• Creator: a daily brief of comment spikes, view drops and sponsor emails makes your first hour about decisions, not scrolling.

Still ahead on the Gemini Edition: a governed-agent design, a way to split big research jobs, and a prompt you can paste. Behind this show is a daily scan of AI news articles, so you don't have to read them.

AWS capacity reportedly sold out through 2027; DeepSeek and MiniMax register for overseas expansion

A Chinese morning tech briefing (mrjjxw) reports two things: DeepSeek and MiniMax, Chinese AI model companies, completed registration for overseas expansion of their large models; and AWS, Amazon's cloud arm, has computing capacity sold through 2027 with large amounts of 2028 capacity booked in advance. Status: REPORTED, single outlet. What the registration involves, and which markets it covers, is STILL OPEN. Plain comparison: through 2027 is roughly the next fourteen months, our own estimate from today. Why it matters: scarcity can shape everyone's bills and limits, but today's stories say nothing about Google's own capacity, so any effect on the tools you use is inference. The expansion item points the other way: more model choices may reach more countries. Practical takeaway: keep prompts short and workflows portable between providers, so a price change or usage cap doesn't strand you, and check licensing and data handling before adopting any newly available model. Prices: not stated. Source: mrjjxw

Anthropic research: AI still can't do genuine real-world science

36Kr reports Anthropic (maker of Claude, an AI assistant) has research confirming AI still cannot conduct genuine real-world scientific research. Status: REPORTED, secondhand; how it was measured and what 'genuine' means are STILL OPEN. Our reading, not the paper's: watching a business for anomalies is a different job from running experiments. A monitor compares today with yesterday inside data you already have. Real-world research means choosing a question, handling messy physical or human conditions, and deciding what the result means. Why it matters for you: it is a useful brake on hype. If you use an assistant to summarize studies or suggest hypotheses, treat the output as a draft to check. The lesson that travels: agents draft and watch well; people still own conclusions. Price: not stated. Source: 36Kr

'Brain rot' for AI: content-farm text may degrade reasoning

technews.tw reports a study warning that absorbing too much content-farm text, meaning low-quality, mass-produced web writing, degrades an AI model's reasoning, and that the damage is hard to reverse. Status: REPORTED, secondhand. Which models were tested, how much junk text caused the drop and how reasoning was scored are STILL OPEN. Why it matters: most people will never train a model, but many feed assistants documents, scraped pages or saved links. Our inference, not the study's: the same logic argues for caution about what you hand an assistant as source material. Practical takeaway: curate inputs. Give it your own notes, official documents and a few trusted pages rather than a pile of search results. If you do fine-tune or build a knowledge base for a team, filter low-quality sources before they go in, because undoing the effect afterward is reportedly difficult. Price: not stated. Source: technews.tw

Architecting governed, self-evolving AI agents on Google Cloud

The Google Developers Blog has a post titled 'Architecting Governed, Self-Evolving AI Agents on Google Cloud.' Status: PUBLISHED by Google; we are working from the title and the setup in our feed, so the specific components, services and code are STILL OPEN. Our plain reading of the title, not a claim about the contents: 'self-evolving' suggests an agent whose behavior improves over time, and 'governed' suggests rules, oversight and records that keep those changes in check. Why it matters: an agent that quietly changes its own behavior is only useful if someone can see what changed and say no. That is a sensible question to ask of any agent product, whether you build it or buy it. Practical takeaway: before you let an agent run unattended, decide who approves its changes, where its actions are logged, and how you roll it back. Price: not stated. Source: Google Developers Blog

A tour of google/adk-recipes: three layers, one repo

Another Google Developers Blog post, 'Three layers, one repo: a tour of google/adk-recipes,' walks through a code repository. Status: PUBLISHED; what the three layers are, which languages it uses and how mature the examples are is STILL OPEN from what we have. We read ADK as Google's Agent Development Kit, which is our assumption from the repo name. A recipe collection matters because most people learn agent building faster from working examples than from documentation. Why it matters for you: if you are a student or job seeker, a repo of tested patterns is a faster route to a portfolio piece, like the job-posting monitor above, than starting from a blank file. Practical takeaway: read the tour first, pick the smallest recipe, run it unchanged, then alter one thing. Check the repo's license before reusing code in client or employer work. Price: not stated. Source: Google Developers Blog

A-MapReduce: splitting wide searches across agents

An arXiv paper, 'A-MapReduce: Executing Wide Search via Agentic MapReduce,' proposes a way to handle 'wide search' tasks. Status: PREPRINT, not peer-reviewed as far as we can tell; its results, benchmarks and costs are STILL OPEN here. Plain explanation, our reading of the title: wide search means gathering many items across many sources, such as every competitor's pricing. MapReduce is an older idea of splitting a big job into small pieces that run side by side, then combining the answers. Applying it to agents suggests many small agents each take one piece. Why it matters: long research requests often fail because one assistant tries to hold everything in a single conversation. Practical takeaway you can try today without the paper: break the job into one-item questions, run each separately, then ask for a combined table. Price: not stated. Source: arxiv.org

Ontology-guided knowledge graphs from climate science papers

A second arXiv paper, 'Ontology-Guided Knowledge Graph Construction from Structured Metadata and Climate Science Publications,' covers building a knowledge graph from structured metadata and climate science publications. Status: PREPRINT; accuracy, scale and which tools it uses are STILL OPEN. Plain explanation: a knowledge graph is a network of linked facts, such as 'this study measured that variable.' An ontology is an agreed list of categories and relationships that guides what gets extracted. Why it matters beyond climate: the title implies the structure is decided first and extraction follows it, which tends to give more consistent results than asking an assistant to 'pull out the key points.' Practical takeaway: if you organize research, customer feedback or contracts, write down your five or six categories before you ask an AI to sort anything. Price: not stated. Source: arxiv.org

Project Chrysalis: governed, self-evolving agents on Google Cloud

Google's developer blog describes Project Chrysalis, a reference design combining Gemini Enterprise Agent Platform (Google Cloud's enterprise agent service), the Agent Development Kit, BigQuery (Google's data warehouse) and a four-gate defense-in-depth security stack that includes Model Armor (a Google Cloud safety-screening service) and Cloud KMS (its encryption-key service). CONFIRMED: Google's post says this. STILL OPEN: how it performs in real deployments. Price not stated. Source: Google Developers Blog

Defense in depth means no single lock: four gates in a row, so if one fails the next catches it. Self-evolving, per the title, means the agent changes over time, which is the risk, since an agent that improves itself can also drift. From the summary we cannot say what each gate checks beyond the two named; read the post for that.

Analysis, three questions to ask of any agent design from any vendor: What may the agent change about itself? What approves the change? Where is the record?

Worth it for: security-conscious teams deciding whether agents may act on their own. For a solo creator it is overkill. AWS appears in today's capacity story, but nothing today describes an AWS equivalent, so no comparison is claimable. It also ties to the Anthropic report: if AI cannot yet be trusted with the whole job, governance is how you use it anyway.

For builders: A-MapReduce, wide search as a map-and-reduce job.

The arXiv paper A-MapReduce: Executing Wide Search via Agentic MapReduce is a posted revision, not necessarily peer reviewed. We have only the start of its abstract, which says LLM-based multi-agent systems (teams of AI agents) show advantages in deep research tasks that emphasize iterative reasoning. Its results are STILL OPEN. Price not stated; worth reading for builders only. Source: arXiv

MapReduce is a classic big-data pattern. Map: split a huge job into small, independent pieces run in parallel. Reduce: merge the partial answers. Counting words across a million documents: each worker counts its own pile, then totals are added.

Wide search, in our gloss, means breadth: many threads, same question, such as filling a table of facts for sixty companies. Analysis of how the pattern would apply: a coordinator splits the list; sub-agents each research one chunk with a fresh, short context; a reducer merges, removes duplicates and checks format.

Why it could work: short context leaves less to forget; a failure stays inside one chunk; chunks run in parallel so you wait less.

Costs and risks: more calls likely means more spend; each sub-agent can be wrong independently; merging is where duplicates and inconsistent formats appear, so the reduce step must verify, not just stack.

Novel or incremental? The pattern is decades old; what is claimed new is applying it to agents. Until reproduced, treat it as an interesting design, not proof.

No code needed: split sixty items into four chats of fifteen with identical instructions, then use one last chat to merge and flag conflicts.

The exceptions-only morning brief. Illustrative workflow, not tested by us. Who it is for: anyone with a daily numbers routine. Price: not stated.

1. Pick five numbers you already check each morning: sales, refunds, support tickets, ad spend, cancellations.
2. Beside each, write in numbers what normal looks like and what counts as an alert.
3. In the Gemini app, create a Gem (a saved custom assistant), or in Google AI Studio (Google's developer playground) set a system instruction (a standing instruction). Paste the prompt in the next section. ChatGPT or Claude can do the same with a saved instruction.
4. Each morning, paste or upload yesterday's data export and ask for exceptions only.
5. Verify: for the first week, check every alert against the raw data, and on one day check by hand for something it missed. Misses matter more than false alarms.
6. Only after a week of good results, consider automating the daily run.

Mistake to avoid: trusting a report you never spot-check. Privacy: do not paste customer data into a tool your company has not approved.

Limits: our source does not describe Pichai's setup, so this is a small-scale version of the idea, and it is only as good as your thresholds.

About 25 minutes a day, roughly two hours a week (our estimate, not a measurement). Assumption: you spend about 30 minutes sweeping dashboards, and the exceptions brief cuts that to about 5 minutes of review.

Options side by side: Gemini (Google), ChatGPT (OpenAI) and Claude (Anthropic) can each take a pasted table; plain spreadsheet rules in Google Sheets or Microsoft Excel can handle the simplest threshold checks with no AI at all. No prices are stated in today's stories, so check the current price on each pricing page.

What you give up: a human glance. The AI may miss a slow drift no rule describes, so keep one full manual review a week.

With AWS capacity reportedly sold out into 2027, a short, portable prompt is cheap insurance.

Prompt clinic: the standing instruction for your morning brief. Illustrative, not tested.

Weak: 'Summarize my sales data and tell me what's important.' It has no thresholds, no format and no permission to say nothing happened, so you get a summary of everything.

Better:

You are my morning watch assistant. I will paste yesterday's numbers.
My normal ranges: [metric 1: range], [metric 2: range], [metric 3: range].
Report only metrics outside range, or moving more than [X] percent versus the prior day.
For each, give: the metric, the value, the normal range, the exact row or cell you used, and one possible cause labeled as a guess.
If nothing is outside range, say 'No exceptions' and stop.
Never invent a number. If data is missing, say what is missing.

Why it works: thresholds define an alert; the cell reference lets you check it; the guess label separates fact from speculation; 'No exceptions' removes the pressure to invent something.

Customize: swap the metrics for job postings, comments or ad spend. Quality check: each morning pick two alerts and match them to the raw data yourself.

What it found: technews.tw reports a Texas A&M study warning that when a large language model (the kind of AI behind chatbots) keeps absorbing popular but low-quality social media text, its reasoning can degrade, like 'brain rot,' and the damage is hard to reverse. Status: REPORTED. We saw only the news summary, not the paper, so method and size of the effect are STILL OPEN. Source

Decision this week: if you feed an AI your own material, through fine-tuning (extra training) or a folder of documents it reads, choose by quality, not popularity. Keep a small clean reference set and re-run the same five test questions after adding material. If answers get worse, remove what you added.

Caution: the study concerns training; a document folder is a looser analogy, so treat this as cheap caution, not a proven fix.

Analysis, not fact: Agentic AI led today's coverage, with open-source AI, policy and China AI also featuring prominently. Yet today's agent stories arrive paired with limits: capacity shortages, governance, data quality and research gaps.

What to watch: official launches from DeepSeek and MiniMax (would confirm the overseas-expansion reports) and any cloud provider confirming its capacity position (would confirm or challenge the AWS report).

I told my AI agent to watch the company 24/7. It filed a report on my lunch break.

Why a few minutes a day is worth it: learning with a reason, from cost savings to which tools to use to real use cases from around the world, today including China's DeepSeek and MiniMax going overseas.

That's it for today. Tomorrow we're watching whether DeepSeek and MiniMax's registrations turn into actual overseas launches, and whether Anthropic's research appears in a form we can read directly.

If this made you smarter, share it with one person who needs it. Thank you for spending your time with us.

Made by AI.