SI Agent News

• Claude Opus 5.5 and video (REPORTED): Jasmine News says the model, described as a general-purpose language model, pieced together a promotional video from a text description alone. Details are thin and we could not confirm the model name from Anthropic in today's material. Jasmine News
• Remote MCP for circuit-board design (REPORTED): a dev.to write-up describes prompting from a MacBook Air while Cadence Allegro X, a hardware design suite, executes on a Windows machine through a remote MCP (Model Context Protocol, the standard that connects AI models to outside tools) setup. dev.to

Claude's 'self-evolution' timeline has reportedly moved up to 2027. Phoenix New Media's tech channel, ifeng Tech, ran the headline 'AI starts building the next generation of AI'. It says the timeline for Claude's 'self-evolution', meaning AI taking a real part in building its own successors, was brought forward to 2027. The story matters because it describes a shift in who does the work of improving AI, not just a faster product cycle. If models help build their successors, the gap between releases could shrink and tools could change under you more often. ifeng Tech

Status: REPORTED. We only have the headline and a short summary. It does not say who set the date, and we found nothing from Anthropic in today's material confirming it, so the date itself is STILL OPEN. Treat it as a forecast, not a schedule.

What this means for you:

• Small business owner: do not plan hiring or tooling around a 2027 date; keep your AI tools swappable.
• Job seeker or student: get good at reviewing and directing AI-written work; that skill pays off whichever date turns out right.
• Creator: expect tools to change faster, so build your process around results you can check, not one tool's features.

Anthropic's IPO valuation reportedly slips to $1.5T from $2T (REPORTED). The site shattered.io carries the headline 'Anthropic IPO Valuation Slips to $1.5T From $2T [2026]'. From the headline alone, an expected valuation for a possible Anthropic listing has come down by about a quarter. We have no filing, no statement from Anthropic and no detail on who produced either number, so treat both figures as market chatter and the underlying facts as STILL OPEN. For most readers this is not a trading signal. It is a reminder that the company behind Claude is a business with investors to answer to, and that pricing, usage limits and product priorities can shift as that pressure changes. If Claude sits at the core of your workflow, keep a second provider tested and ready, and do not read a valuation headline as a verdict on the product's quality. Source: shattered.io.

Anthropic pre-IPO futures start trading on Hyperliquid with up to 10x leverage (REPORTED). Bloomingbit reports that futures tied to Anthropic's expected IPO have begun trading on Hyperliquid, a crypto trading venue, with leverage of up to 10x. We only have the headline, so details such as contract terms, volumes and who is behind the market are STILL OPEN. Two general points are worth knowing. A futures contract is a bet on a price, not ownership of Anthropic shares. And with 10x leverage, a move of roughly 10% against a position can wipe it out, before fees. The practical takeaway for non-traders is to treat any price printed from such a market as a thin, speculative signal and not as a fair valuation. It is also a useful cross-check on the valuation story above: both are unconfirmed, and neither should drive a business decision. Source: bloomingbit.

Claude Code restrictions on game ports (REPORTED). TweakTown reports Anthropic is going after vibe-coded game ports, meaning games moved to a new platform using AI-written code, with new Claude Code restrictions. The summary does not say what the restrictions are, so details are STILL OPEN. The wider point is that the maker of a coding tool is drawing lines around what it will help produce, especially when the code reproduces someone else's finished product. If you port, remake or rebuild existing software with Claude Code, read Anthropic's current usage policy and the Claude Code documentation first, and keep records showing you have the rights to the original. A hobbyist can lose a project's momentum to a refused request, and a studio can face bigger problems if a workflow depends on something the tool no longer allows. TweakTown

DeepSeek's agent architecture (REPORTED). Ifanr reports DeepSeek, an AI lab, released a new agent architecture in which AI directs AI to do the work, and frames humans as the next bottleneck for agents. We have only the summary: no benchmarks, no pricing. In plain terms, one AI plans and hands pieces of a job to other AI workers, so a person is no longer the one splitting up and passing along each step. That can speed up long tasks, but it also means errors can compound before anyone looks. The practical question for Claude users is where your human checkpoint sits: after the plan, before anything is sent or published, or only at the end. Decide that deliberately for any multi-step automation, especially one that touches customers, money or published work. Ifanr

Claude Opus 5.5 reportedly assembles a promo film from a description alone (REPORTED). Jasmine News runs the headline 'General-purpose large model crosses into video: Claude Opus 5.5 assembles a promo film from a description alone'. We only have that headline and no demonstration, sample output or specification. We also could not confirm the model name or the claimed capability from any other source in today's material, so the details are STILL OPEN. If it holds up, the notable part is that a text-first model is being used to put together video without a separate video tool. Creators and small businesses that need short promos could save real time. Before relying on it, run a small test: give it your product description, check the result for factual errors and brand fit, and note how much editing is left. Judge it by that test, not by a headline. Source: Jasmine News.

Vedha: testing a web app with an AI agent in five minutes (REPORTED). A dev.to post introduces Vedha, a testing framework for what it calls the agentic world. The headline says you can test your web app with an AI agent in five minutes and catch bugs that something would miss; the rest of the title is cut off in our material, so what it compares itself with is STILL OPEN. The idea is useful regardless: instead of writing scripted tests, you let an AI agent use your site as a person would and report what breaks. That suits solo builders and small teams with no dedicated testers, and it matters more as AI writes more of the code, because AI-written code still needs checking. Try it on a staging copy, not your live site, and treat its findings as leads to verify. Source: dev.to.

Tool selection at scale: what Top-K RAG leaves out (REPORTED). A dev.to post, 'Tool Selection at Scale: What Top-K RAG Leaves Out', looks at a problem that appears once an AI agent has many tools. A common approach is to search the tool descriptions and show the model only the K closest matches. We only have the title, so the article's own findings are STILL OPEN. The weakness the title points to is easy to reason about: if the right tool is not among the top K matches, the model never sees it and cannot choose it, however capable it is. For anyone building agents or connecting many MCP servers, the practical lesson is to test with realistic requests, check whether the needed tool was actually offered, and avoid assuming that more tools means more capability. Group tools sensibly and keep descriptions clear and distinct. Source: dev.to.

Prompting on a MacBook Air, executing on Windows with remote MCP (REPORTED). Another dev.to post describes building remote MCP infrastructure for Cadence Allegro X, a circuit board design program. The title sets out the idea: you write prompts on a MacBook Air while the work runs on a Windows machine. MCP is the standard that lets an AI assistant call outside tools, and running it remotely lets a light laptop control software that only runs elsewhere. We only have the title, so the setup, security approach and results are STILL OPEN. The pattern applies well beyond circuit design: any Windows-only or heavy desktop program could be driven this way. If you try something similar, think first about access, since a remote tool server can act on your files and software. Limit what it can do, require authentication, and keep it off the open internet. Source: dev.to.

Anthropic's valuation slips on paper, and a market for betting on it appears. shattered.io reports Anthropic's IPO valuation slipping to $1.5 trillion from $2 trillion. That is $0.5 trillion lower, or a quarter off the earlier figure. shattered.io Separately, bloomingbit reports pre-IPO futures on Anthropic have begun trading on Hyperliquid, a trading platform, with leverage of up to 10x. bloomingbit

Plain terms: a pre-IPO future is a contract tied to what a company might be worth before its shares trade publicly. Ten-times leverage means a position ten times your cash, so a move of roughly 10% against you can wipe out the stake, before fees. This is not investment advice.

Status: both reports are REPORTED. We could not verify either from a primary source today. Whether an IPO happens, when, and at what price is STILL OPEN.

Analysis (ours): a valuation is investors' opinion of future value. It is not the price, quality or availability of the model you use. For builders on Claude the useful reading is vendor risk, not market timing. Keep prompts and tool definitions in your own files, log which model handles which task, and keep one tested fallback model. Then a price change, policy shift or outage is an afternoon's switch, not a rewrite.

For builders: a dev.to write-up, 'Tool Selection at Scale: What Top-K RAG Leaves Out', says the usual recipe (embed, retrieve, let the LLM decide) works for about 100 tools, but at 1,000 you need a 'control plane'. dev.to This is a practitioner post, not peer-reviewed, and we have not run it. Status: REPORTED.

The simple recipe: turn each tool's description into an embedding (a list of numbers that captures meaning), find the few closest to the user's request, and hand the model that shortlist.

Why it breaks (our explanation): at 1,000 tools many descriptions look alike ('create ticket', 'open ticket', 'new issue'), so similarity search cannot separate them and the right tool falls out of the top few.

The fix, in four parts of the post's five:

• Hybrid retrieval: combine meaning-based search with plain keyword matching, so an exact tool name still hits.
• Hard filters: rules that remove tools before any ranking, such as tools the user may not use or the wrong environment. They are deterministic, so they cannot be fooled.
• Learned reranker: a second model that rescores the shortlist by reading the request and each description together, trained on examples of right choices.
• Progressive disclosure: show the main model only names and one-line descriptions first, and load full definitions only for the tools it picks.

The fifth, per-stage evals, is in the research note below.

Novel or incremental? Incremental. Search teams have used hybrid retrieval and rerankers for years. What is new is treating tool choice as a testable pipeline.

Why Claude users care: every MCP server you connect adds tools. A few installs can pass 100 without you noticing. Symptoms: near-miss tool calls, or context eaten by definitions.

Lesson, not news: make your agent prove the outcome, not just the screen.

A dev.to post introduces Vedha, a testing framework where an AI agent tests your web app, billed as a five-minute test. Its pitch: an agent sees 'Order placed' on screen and reports success even when the API behind it returned an error. dev.to Status: REPORTED. We have not tried Vedha; price is not stated. Worth a look if you ship a web app with checkout or signup and no automated end-to-end tests.

Who it's for: developers using Claude Code on a web app. Illustrative workflow, not tested:

1. Pick one flow that costs money if it breaks, such as checkout or signup.
2. Define 'done' as two facts: what the screen says and what the backend says (response status code, a row in a test database, or a queued email).
3. Give Claude Code a way to see the backend: server logs, a test database it can query, or the browser's network log.
4. Put the rule in your project's CLAUDE.md file, which Claude Code reads for project instructions.
5. Run it, then break something on purpose, such as making the test API return an error. If the agent still reports success, the rule is not wired in.

## Flow checks
A flow passes only if the screen result and the backend result agree.
For every step, report: screen says / backend says / verdict.
Any mismatch is a FAILURE, even if the screen shows success.
Use the test environment only. Never trigger real payments.

Expected output: a short table per step. Limits: needs a test environment, and the agent can still misread logs, so spot-check the first few runs by hand. Verify it worked: step five, the deliberate failure.

Lesson, not news: send a shortlist of tools, not all of them.

Every tool definition you send rides along in every request, and you pay for those input tokens (the chunks of text models bill by) each time. Today's tool-selection write-up gives the numbers to work with: 1,000 tools versus a shortlist. Say you shortlist 20. That is 50 times fewer definitions per request.

Illustrative estimate, not a measured result: assuming about 100 tokens per definition (replace with your own count), that is roughly 100,000 input tokens per request down to about 2,000.

How to measure: most model APIs report input token counts per request, so log the number before and after. Anthropic, OpenAI and Google all bill by tokens; check the current price on each company's pricing page. We are not quoting prices from memory.

What you give up: if the right tool is not in the shortlist, the agent cannot choose it, and building the shortlist takes engineering time. Mitigation: if the agent says nothing fits, widen the shortlist once. Worth it if you have well over 100 tools; below that, probably not.

Prompt clinic: rewrite your tool descriptions so lookalike tools stop colliding.

Weak prompt: Write descriptions for my tools. You get flattering blurbs that all sound alike, which is exactly what makes tools collide.

Better prompt (copy-ready):

I'm giving you a list of tools my AI agent can call. For each tool, write a description of at most 40 words using this pattern: (1) what it does, in plain verbs; (2) when to use it; (3) when NOT to use it, naming the closest similar tool in my list; (4) the exact names of its inputs. Avoid adjectives like powerful or seamless. Then list any pairs of tools whose descriptions still overlap, and suggest which one to rename or merge. Here are the tools: [paste names and current descriptions]

Why it works (our reasoning, not tested): the 'when NOT to use' line separates lookalikes, the overlap list surfaces collisions, and banning adjectives keeps the keywords that retrieval matches on.

Expected result: tight descriptions plus a short collision list. Customize: add your domain's terms and any naming rules. Quality check: take ten real requests and see whether the right tool lands in the top five, before and after.

Not a study, an engineering lesson: the same tool-selection write-up advises per-stage evals, meaning you test each step of the pipeline on its own. An eval is a test set with known right answers. It is a practitioner post, not peer-reviewed.

Why it matters: if the right tool never reaches the shortlist, no model can pick it, so measure shortlist recall, the share of requests where the right tool is in the shortlist at all.

Decision this week: write ten real requests with the correct tool noted, run only your retrieval step, and count hits. As our own rule of thumb, if fewer than nine land, fix retrieval before touching prompts.

Agents are the busiest topic area today, and several reports, including the DeepSeek and self-evolution stories, come from Chinese tech press, so reading across regions pays off. Three things to watch, none with a firm date:

• STILL OPEN: whether Anthropic itself says anything about a 2027 self-evolution timeline. An official post would confirm it; silence keeps it a claim.
• STILL OPEN: whether a formal IPO filing appears. A public filing would confirm; more price chatter would not.
• STILL OPEN: what the Claude Code restrictions actually say. Published usage policy or documentation would settle it.

My agent reported checkout success. It had only verified the confetti animation.

Learning every day with a reason: cost cuts like today's shortlist trick, which models to use, and real use cases from around the world, including China in today's DeepSeek report. Thanks for spending part of your day with us. If this made you smarter, share it with one person who would find it useful.

SI Agent News — the world's first podcast newsletter. Made by AI.