SI Agent News

• Yahoo's coverage of Claude in science says Claude Code helped scientists improve cancer therapy components. REPORTED, and the headline does not say what was measured against what. Yahoo!
• Off-topic flag: a search for 'Claude' today also surfaces an NHL postgame clip about the hockey player Claude Giroux. It is not about the AI. NHL.com

A biotech startup says Claude ran a mouse brain surgery on its own. Exin Therapeutics, a US biotech startup, claims it let Anthropic's Claude (an AI model) make autonomous decisions and carry out gene-therapy brain surgery in mice, as Tech News Taiwan reports. Status: REPORTED, and the claim comes from the company. Oversight, success rate and what a person approved along the way are STILL OPEN, and we have not seen independent confirmation. What is new is the shape: not a chatbot answering questions, but a model choosing an action and a system carrying it out in a physical lab. That is the same pattern you will meet in everyday software, where an agent clicks, sends or buys instead of suggesting. Price not stated; this is a research claim, not a product. The useful habit is to ask three questions of any agent you let act: what can it touch, who approves the risky steps, and where is the log. technews.tw

What this means for you:

• Small business owner: if an agent will ever book, order or email for you, decide now which actions need your sign-off.
• Job seeker or student: lab and operations teams will want people who can supervise an agent, so practice writing the rules one must follow.
• Creator: an agent that acts on your accounts can publish or delete, so start it read-only.

Anthropic reportedly plans a $3.5 billion Japanese share offering ahead of a Nasdaq listing. 36Kr reports the plan, which would put a large block of shares in front of Japanese investors before a US listing. Status: REPORTED; the timing, pricing, who can buy and whether the listing date is fixed are STILL OPEN, and the summary we have does not say. Why it matters to you: a company that raises money this way is signalling it expects to be around and funding heavy compute for years, which bears on whether you build a workflow, a product or a client service on its tools. It is not a reason to buy shares. Pre-listing stories shift quickly, and an offering announced in one market can change before it closes. If you depend on Claude for work, the practical step is unrelated to the stock: keep your prompts, files and processes portable so a change in price or terms does not strand you. This is information, not investment advice. 36Kr

Anthropic may be rated junk, and pre-IPO perpetual futures fell 20% in a month. Finance.technews.tw reports that Anthropic could receive a junk (below investment grade) credit rating and that perpetual futures tied to its pre-IPO value dropped about 20% over a month. A perpetual future is a derivative with no expiry that tracks a price, here a company that is not yet publicly traded, so these markets can be thin and jumpy. Status: REPORTED as market chatter; whether any rating agency has actually issued a rating, and what is driving the drop, are STILL OPEN. Our reading is that a 20% move in a thin, speculative product says more about trader mood than about the company's finances. For a reader who uses Claude, the takeaway is modest: do not treat a headline price move as a sign the product will change next week, and do not trade on it without understanding the instrument. Again, not investment advice. finance.technews.tw

Claude reportedly cut pharma document review time by 89%. TechRadar reports an 89% cut in pharmaceutical document review times, roughly the gap between an eight-hour day and about an hour (our own rough estimate from their figure). The baseline and scope are not in the summary, so treat it as REPORTED. The coverage says Anthropic believes the way to find bottlenecks is to get deeply involved in real drug programs; our reading is that the gains came from embedding in a specific process rather than from a general trick. The transferable lesson is to pick one long, repetitive document task, time it by hand first, then measure again with the AI and keep a human check on the output. Without a baseline, any percentage is hard to trust, including this one. Worth it for anyone buried in long, repetitive documents; price not stated. TechRadar

Update: Claude Code's video-game limit draws false-positive complaints. Anthropic has limited Claude Code (its coding tool) from recreating specific existing video games. International Business Times UK reports independent developers hitting refusals on projects they say are legitimate, including commissioned work. The limit looks deliberate (REPORTED); whether Anthropic adjusts it is STILL OPEN. The risk for you is a mid-project block when a request resembles an existing title, such as a game in a familiar genre or a tribute. Our advice, not Anthropic's: describe your project's design in your own words, avoid naming the original game in prompts, and keep commission paperwork handy in case you need to explain the work. Also keep your code in version control so a refusal never costs you progress. IBTimes UK · via MSN

New paper asks how much your tool calls give away. An arXiv paper titled "How Much Can Tool Calls Reveal About You? Privacy Leakage from Tool-Call Traces" examines what the record of an AI agent's tool use can expose about the person behind it. We have only the title and framing, not the findings, so the results are STILL OPEN to us and it is a preprint, meaning not yet peer reviewed. The idea is easy to grasp: when an agent searches, opens files or calls a calendar, the search terms, file names and timing form a trail, even if your chat text is never shared. That trail may sit in logs held by the tool provider, a plugin or your own company. Practical steps: check which tools your agent connects to, read what each logs, avoid putting personal details in tool arguments when you can, and favour connections you can switch off. arxiv.org

Your AI agent remembers things, but who audits the memory? A dev.to post raises a question that grows as agents keep notes between sessions: who checks what the agent has stored? We have the headline and its framing, so details of the author's proposals are STILL OPEN. It matters because a wrong or outdated stored note can quietly shape later answers, and a stored private detail can resurface where it should not. A simple routine works for non-developers too: find where your tool keeps memory, read it monthly, delete anything stale or sensitive, and test by asking what the agent knows about you. If you build agents, log when a memory is written, from what source, and let users see and edit it. dev.to

Not our beat: a hockey interview that matched on the name "Claude." NHL.com published a postgame media session with Claude Giroux after a game against Anaheim. It concerns the hockey player and has no connection to Anthropic's Claude, so we are not treating it as AI news. Status: not relevant, and we are not drawing anything from it. We mention it because name collisions like this are a small, practical hazard: news alerts, social feeds and automated summaries can mix up a person, a product and a company that share a word. If you follow AI through keyword alerts, add the company name (Anthropic) or the product name, and treat anything lacking that context with suspicion before you share it. NHL.com

Anthropic's IPO: a $3.5 billion Japanese offering and some market jitters.

36Kr, citing Cailianshe, reports documents showing Anthropic plans a $3.5 billion Japanese stock offering before listing on Nasdaq. REPORTED; not confirmed by the company in anything we have. 36Kr

Separately, Tech News Taiwan's finance site reports that pre-IPO perpetual futures (contracts that let traders bet on a price with no expiry date) fell 20% in a month, and that Anthropic may not secure an investment-grade credit rating after listing. In plain terms, a contract worth $100 a month ago is now about $80 (our illustration). A rating is a forecast, not a fact, so the junk-grade question is STILL OPEN. finance.technews.tw

Why a Claude user should care (our analysis, not fact):

• A public company discloses more about its costs and risks, which could tell you how stable your vendor is.
• A weaker credit view can make borrowing more expensive.
• A vendor under market pressure might change prices or plans. Nothing in today's stories says that will happen.

Who benefits and who is exposed: early investors and buyers of the offering are the obvious stakeholders; customers building on Claude carry vendor risk if they have only one provider.

What to do: do not panic-switch. Keep prompts portable, log your monthly Claude spend, and know which two or three workflows you could move to another provider within a week. That is resilience, not pessimism.

For builders: how agent memory actually works, and why it needs an audit trail.

A dev.to opinion piece notes that memory tools such as mem0, Zep, Letta and Cognee store what your agent 'learned', and asks who wrote each fact. It is commentary, not news (REPORTED as an argument, not a finding), and today's material does not describe those products' internals. What follows is the general pattern, labelled as our analysis. dev.to

The mechanism:

• Write path: after a conversation, the system extracts statements such as 'this customer wants invoices in euros' and saves them.
• Store: a database holding the text plus an index for searching.
• Read path: next session, relevant memories are retrieved and placed into the prompt before the model answers. The model remembers nothing; it reads notes. Memory is prompt text you did not write.

Three failure modes: a wrong fact the model misread and now repeats every session; a poisoned fact, where text from a web page or email the agent read is saved as if it were your preference; and a stale fact, such as an old job title.

A proposed audit (our analysis): every memory record carries four fields: the text, its source, when it was written, and whether a person approved it. Review new memories weekly, delete by source, and hold anything learned from untrusted content for approval. Routine facts you state directly can save automatically; reads stay instant.

Novel or incremental? Incremental. Nothing here is a new technique; the gap is governance. For a toy assistant the overhead is heavy. For an agent touching customer data or money, the audit trail is how you explain a wrong action afterward.

The autonomy ladder for Claude Code. This is a lesson inspired by today's agent stories, not news; the steps are illustrative, not tested by us.

Who it is for: anyone letting an agent edit real code or files. Cost: Claude Code pricing is not stated in today's stories; check Anthropic's pricing page.

1. List what your agent can do and sort it into three tiers. Free: read files, run tests. Ask first: edit files, install packages, call outside services. Never: delete data, push to your main branch, touch production keys.
2. Write the tiers, one plain sentence per rule, into a file called CLAUDE.md in your project's root folder. Claude Code reads that file at the start of a session.
3. Start each task with: 'Do not edit anything. Give me a numbered plan, the files you would touch, and the risks.'
4. Trim the plan, then approve in slices: 'Execute steps one to three only, then stop and show me the diff.'
5. Run the tests yourself and ask it to explain each change.
6. Have it append what it did, with the date, to a log file.

Example: an agent maintaining a website. Free: read pages, run the build. Ask first: edit templates, add a plugin. Never: delete a page, change DNS, publish without you.

Verify it worked: ask for something from the never tier and check that it refuses or asks.

Limits: a file of instructions is guidance, not enforcement. For hard limits use read-only credentials and your tool's permission settings. Mistake to avoid: giving it write access to production 'just for now'.

Put a number on the pharma result, with a haircut. The 89% cut in document review time is a pharma figure (TechRadar), so do not plan around it. Assume you spend four hours a week on first-pass reading of contracts or reports. If an assistant saves even a quarter of that, it is one hour a week, about 50 hours over a year of 50 working weeks. That is our estimate with those assumptions, not a promise.

Test two, pay for one. Claude (Anthropic's assistant), ChatGPT (OpenAI's assistant) and Gemini (Google's assistant) can all read documents. Run the same three of your documents through two of them, time it, and count the errors. Check each company's pricing page for current prices; none is stated in today's stories.

Second minute-saver: ask for the table in today's prompt, then paste only the flagged rows into a follow-up instead of the whole document. Shorter inputs mean faster answers and, on pay-per-use plans, usually fewer tokens (the chunks of text models bill by).

What you give up: accuracy you must verify, since a fast wrong summary costs more than a slow right one, and privacy: read each tool's data terms before uploading anything confidential. Spend the saved hour on checking, not on more documents.

Prompt clinic: document first pass. Illustrative, not tested by us.

Weak: 'Summarize this contract.' You get a smooth paragraph that may skip the clause that matters, and you cannot tell what was invented.

Better:

You are helping me with a first-pass review of the attached document. This is not final advice.
1. Give a five-line plain-language summary.
2. List every deadline, payment amount and obligation, each with the exact quote and its section or page.
3. Flag anything unusual, ambiguous or missing, and say why.
4. If you are unsure about anything, write UNSURE instead of guessing.
5. Do not add information that is not in the document.
Put the results in a table with four columns: item | exact quote | location | confidence.

Why it works: it scopes the job as a first pass, demands exact quotes so you can check, permits 'UNSURE' which cuts guessing, and forces structure.

Customize: add your role and what you care about, such as renewal dates or liability.

Quality check: pick three rows at random and find each quote in the original. If one is missing or reworded, rerun before trusting the table.

One more pair, for Claude Code. Weak: 'Fix the login bug.' Better: 'Login fails after a password reset. Do not edit anything yet. Find the likely cause, list the files involved, and propose a fix with a test that fails today and passes after.' It states the symptom, separates diagnosis from change, and defines done.

A new arXiv paper (the open research archive) is titled How Much Can Tool Calls Reveal About You? Privacy Leakage from Tool-Call Traces. Definitions: a tool call is an AI agent asking an outside tool to do something; MCP, the Model Context Protocol, is a standard way to connect agents to tools; a trace is the record of every call and its inputs. We only have the abstract's opening, which says agents use such protocols and produce traces, so we quote no results. What the paper shows is STILL OPEN; read it before citing it.

Decision this week: treat tool-call logs as personal data. Find where they are stored, who can read them, how long they are kept, and whether they contain names, addresses or search queries. arxiv.org

Agentic AI is one of the busiest topics today, alongside open-source models, Claude, Chinese AI and policy. Claude's items cluster around agents doing real work, the rules around them, and money.

Watch, with no announced dates (expectations, not scheduled events): (1) independent confirmation of the mouse surgery from anyone besides Exin, which would turn a company claim into a result; (2) official details on the reported Japanese offering and any credit rating; (3) whether Claude Code's game limit is adjusted after the false-positive reports.

I gave my agent memory; now it remembers every time I said 'just this once.'

That's it for today. Tomorrow we're watching whether anyone independently confirms the mouse surgery, and what more surfaces about Anthropic's Japanese share offering. If this made you smarter, share it with one person who needs it. Thank you for reading and listening.

SI Agent News — a podcast newsletter. Made by AI.