SI Agent News

• ChatGPT Business admin control (CONFIRMED by OpenAI's community forum post title; details not read; price not stated): owners can now restrict who sends invitations to owners and admins only. Anthropic competes with OpenAI for team accounts, so it is worth checking for the same setting on whatever assistant your team uses. Worth it for anyone who administers a team plan.

Anthropic reportedly launches Claude Duet, a personal assistant (REPORTED, one source). ThePaper.cn, a Chinese news outlet, frames it as another giant entering personal AI assistants. The material we have is the headline and little else. No features, price, regions or availability are stated, so we won't guess at any of them. Why it matters: a personal assistant is a different product from a coding tool or a chat window. It implies something that stays with you across tasks rather than answering one prompt at a time, so even a single-source report hints at where Anthropic may be heading. If you already use Claude, nothing changes today. The confirmation to wait for is Anthropic's own announcement or documentation, which would settle what Duet does, what it costs and who can use it. Until then, treat it as a signal to watch, not a tool to plan around. Price not stated. Nothing to act on yet except watching.

DeepSeek reportedly completes a roughly 100-billion-yuan round (REPORTED). MyDrivers reports that DeepSeek, a Chinese AI lab, raised about 100 billion yuan. CATL (a battery maker) and Tencent (a tech giant) are named as the largest contributors, and the outlet frames the raise as taking on Anthropic and OpenAI. We have not converted the figure, so check a current exchange rate before quoting it. DeepSeek's own confirmation is not in what we have, which is why this stays labeled as reported. Analysis: if it holds, a direct rival now has deep pockets. That usually means more model releases and pricing pressure, which tends to help buyers. The practical takeaway is to avoid hard-wiring a workflow to a single vendor. Keep prompts and tool definitions portable so you can switch models if price or quality shifts. Watch for a company statement or a filing that confirms the amount and the investors.

Anthropic in court over song lyrics in Claude training (STILL OPEN). AD HOC NEWS reports that the dispute over song lyrics used in Claude's training remains open. No ruling is stated in what we have, and we won't speculate on how it ends. Why it matters: for most individual users this is not a today problem. It does show that the question of what data trained a model is still being argued in court. If you build a customer-facing product on Claude, one email to your vendor asking how training-data claims are handled and what your contract says is cheap, and it puts the answer on paper. If you only use Claude for internal drafting, there is nothing to change. A ruling or a settlement would resolve the matter. Until one appears, treat it as background risk worth a line in your vendor-review notes.

MCP schema versioning: keeping AI tool calls working through API changes (dev.to). This dev.to piece covers a problem that hits anyone who connects a model to tools through MCP (Model Context Protocol). When a tool's schema changes, calls that used to work can quietly start failing. The article's angle is versioning, so changes don't break the agents and clients that depend on the old shape. We only have the headline and summary, so we can't vouch for its specific recommendations. Analysis: this is the most practically useful item in today's edition if you maintain any tool server. The general habits are cheap. Version your schemas explicitly. Prefer additive changes over renaming or removing fields. Test existing clients against a new version before you ship it. Announce deprecations rather than swapping behavior silently. If you only consume other people's MCP tools, check whether they publish version notes before you upgrade.

MemoryTax: The Bitter Lesson of Agent Memory (arxiv.org). This arXiv paper takes on how AI agents store and recall information across steps or sessions. The title echoes the well-known "bitter lesson" idea in AI, usually summarized as general methods beating hand-built tricks. We have only the title and not the paper's findings, so we can't say what it measures or concludes. We also can't say whether its claims survive scrutiny, since arXiv papers are preprints, not peer-reviewed results. Why it matters: memory is where many agent projects get expensive and fragile. "Tax" suggests a cost that gets paid whether or not the memory helps. If you are designing an agent with elaborate memory layers, this is worth reading before you build more. Check the paper's actual results first, then test whether simpler memory holds up against your own tasks.

ChatGPT Business: owners can restrict invitations to owners and admins (community.openai.com). A post on OpenAI's community forum says ChatGPT Business workspace owners can limit who is allowed to send invitations, so only owners and admins can add people. We have not seen OpenAI's official documentation, so check your own admin settings for the exact wording and where the control lives. Price and plan availability are not stated. Why it matters: if you run a team workspace, open invitations let any member add outsiders or extra seats. That can mean surprise costs, or people with access you never reviewed. Restricting invitations is a quick, low-effort control. Someone with the right role should check the setting this week and confirm who currently holds owner and admin roles. If you are not an owner, there is nothing for you to do except ask.

Steering black-box LLM agents with a "pitfall anticipation world model" (arxiv.org). This arXiv preprint proposes a way to guide LLM agents when you cannot see inside or retrain the model, which is what "black-box" means here. The title suggests a separate model that anticipates where an agent is likely to go wrong, so the agent can be steered away from those mistakes. We have only the title, so we can't report results, benchmarks or whether the approach works outside the paper's own tests. Why it matters: most people building on hosted models can change prompts and tools but not the model itself. Methods that improve reliability from the outside are therefore directly relevant. This is research to bookmark, not something to deploy. If you run agents that take real actions, read the full paper and check whether it names code or an evaluation you can reproduce.

Claude Giroux postgame media vs. FLA (NHL.com). This one is not AI news. Claude Giroux is an NHL hockey player, and the item is a postgame media session on NHL.com, apparently against Florida (FLA). It reached today's feed, most likely because of the shared first name with Anthropic's Claude. It has no connection to Anthropic, Claude the model or anything above. We have no quotes, score or details from the session, and we won't guess at any. Why it matters: only as a reminder that keyword-based news feeds pull in false matches. If you read automated digests, expect the occasional name collision and don't act on it. If you follow hockey, NHL.com has the session itself. For everyone else, skip it. Nothing to act on.

MemoryTax: is agent memory being built backwards?

Training data is what a model remembers permanently. What it remembers about your project is a separate problem, and an arXiv preprint called MemoryTax argues the field has approached it backwards.

What the abstract says (CONFIRMED as the paper's framing): memory systems such as Mem0 and Zep (tools that give agents long-term memory) decide what to keep before any question arrives. Coding agents instead keep the raw history and let the model search it. The title borrows 'bitter lesson,' Rich Sutton's argument that general methods plus computing power tend to beat hand-built cleverness. That context is ours, not the paper's.

Why it matters (analysis): pre-selection is a bet. You guess at write time what will matter, summarize, and discard the rest; if the guess is wrong, the detail is gone. Raw history plus search defers the decision until the question is known, at the cost of bigger history and a dependence on search quality. Summaries are cheaper to read, so pre-selection may win on read-time cost. The title suggests the paper argues for a hidden tax on that approach, but we only have the abstract's opening: no numbers, no method.

The pain it describes (REPORTED): a companion dev.to post says closing a long coding session and opening a new one leaves the agent greeting your codebase like a stranger, with the architecture decision you explained twice and the test command that only works in one setup forgotten. It is a community blog, so experience rather than evidence.

Status: whether the paper's claim holds is STILL OPEN; it is an unreviewed preprint.

Practical read: do both. Keep a short, human-edited notes file for decisions that must never be forgotten, and keep full history searchable for everything you did not predict. Vendors selling memory as a pure summarizing layer will need to show they beat searching the transcript.

For builders: why a harmless-looking change breaks MCP tool calls.

The scenario (REPORTED, dev.to community post): an agent that worked yesterday fails today after a change that looks harmless in a pull request: rename location to query, make a field required, or reshape a result object. The model has not changed.

The mechanism: MCP is the Model Context Protocol, an open standard Anthropic introduced for connecting models to tools. Each tool is published with a name, a description and an input schema (a machine-readable list of fields and types). The model reads that definition as part of its context and writes its calls from it, so the schema is both a contract and a prompt.

• Rename: any client, saved example or cached tool list still using the old field sends something the server no longer knows.
• New required field: older callers omit it and validation rejects the call.
• Reshaped result: the model, or code downstream, expects the old shape and misreads the new one.

None of these throws a compile error. They show up as an agent that mysteriously gets worse.

Rules of thumb (our analysis, not tested):

• Additive changes are safe: add optional input fields and extra result fields.
• For a rename, accept both names for a window and log which one callers use.
• Introduce a new required field as optional with a default first.
• For a truly breaking change, publish a new tool name and keep the old one.
• Write error messages that tell the model how to fix its call, because it reads those too.

Trade-off: every extra tool definition takes room in the model's context, so aliases and old versions are not free. Price not stated; check your provider's pricing page. Status: failure patterns REPORTED by one community post; the rules are judgement.

For builders: a golden-calls check for tool schema changes. Illustrative workflow, not tested by us; adapt it to your stack.

1. Pull about ten real tool calls from your logs: tool name, arguments, and what a good result looked like. Save one per line in a file next to your server code.
2. Write a small script that loads the new tool schema and validates each saved call with any JSON Schema validator. Any failure is a breaking change. Estimate: roughly an hour for a first version.
3. Give the model the new tool list plus ten plain-English tasks and check it picks the right tool with valid arguments. Run each task a few times, since models vary.
4. If you rename a field, add an alias and log every call that uses the old name. Remove the alias once the count stays at zero for a stretch you choose, such as two weeks.
5. Run steps 2 and 3 automatically whenever the schema file changes.

Verify it worked: rename a field on a scratch branch and confirm the check fails. If it passes, it is not looking at the right thing.

Limits: ten calls only cover what you have already seen; rare calls slip through.

Estimate: about 13 hours a year, roughly a day and a half of work. Writing a project notes file takes about ten minutes once. If re-explaining your setup costs five minutes a session and you open three sessions a week, that is about 15 minutes a week. These are assumptions, so plug in your own.

Options from different companies: Claude Code reads a CLAUDE.md file, OpenAI's Codex reads AGENTS.md, and Cursor (an AI code editor) has its own rules files. File names are from general knowledge, not today's stories. Prices: check the current price on each pricing page.

What you give up: upkeep. Stale notes mislead the agent, so budget a few minutes a month, and the file occupies context every session. Keep it to the permanent: build command, settled decisions.

Prompt clinic. Weak version: 'Check if my change to the tool is okay.' It doesn't say what okay means, gives no old version, and invites reassurance.

Better version (illustrative, not tested):

You are reviewing a change to an MCP tool definition. Below are the OLD and NEW tool schemas, followed by 5 recorded calls from production.

1. List every difference between OLD and NEW: renamed fields, new required fields, removed fields, changed result shapes.
2. For each difference, say whether a client built against OLD would break, and why.
3. Check each recorded call against NEW and mark PASS or FAIL with the exact validation error.
4. Propose the smallest change that keeps old calls working, such as an alias or an optional field with a default. Do not rewrite the whole schema.

If anything is ambiguous, ask me before assuming.

OLD: [paste]
NEW: [paste]
CALLS: [paste]

Why it works: it defines 'breaks,' supplies evidence, forces a verdict per item, and caps the fix at the smallest change. Quality check: verify one FAIL by hand with a real validator, since the model is reading, not running, your schema.

Steering Black-Box Large Language Model Agents with Pitfall Anticipation World Model (arXiv preprint, not peer reviewed). 'Black-box' means a model you can only call, not retrain. A 'world model' is a component that predicts what happens after an action. The abstract opens by saying frontier agents are proficient at long-horizon reasoning and precise action generation; the rest of that sentence and the results are not in what we have.

One decision this week (illustrative, not the paper's method): if an agent fails late in long tasks, keep a written list of the specific pitfalls it hit and put that list in front of it on the next run.

Agentic AI was one of the busiest topic areas today. Today's stories touched several: a product report (Duet), funding (DeepSeek), policy (the lyrics case) and agent engineering (memory, tool schemas).

Watch: Anthropic itself confirming or denying Duet (an unconfirmed report, no date), DeepSeek confirming the round (REPORTED, no date), and any ruling in the lyrics case (STILL OPEN, no scheduled date stated).

I renamed one tool parameter for clarity. My agent has since renamed me 'the problem.'

That's it for today. Tomorrow we're watching whether Anthropic says anything about Claude Duet and whether DeepSeek confirms its funding round.

Listening every day is learning with a reason: today it was a time-saving notes file, a way to stop agent tools breaking, and a look at how a Chinese lab is being funded. If this made you smarter, share it with one person who needs it. Thank you for spending the time with us.

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