SI Agent News

• Claude on a TV (Ifanr, a Chinese tech outlet): a developer put Claude onto a TV and showed what changed in its system. The piece is framed around 'the era of great reverse engineering.' The details we have are thin, so treat it as a curiosity. Status: REPORTED, one source.
• Dario Amodei profile (The Profile, by Polina Pompliano): background on Anthropic's chief executive. Little new, but useful context on who steers the company. Status: background, no new claims.

Claude Code gets MCP Tasks and Elicitations — a narrow launch

On October 23, 2026, Anthropic shipped support for MCP Tasks and Elicitations in the Claude Code command line (a terminal-based coding tool). MCP, the Model Context Protocol, is an open standard for plugging AI into outside tools. Yahoo reported it and Forkast News carried the same story. The coverage says MCP tools move beyond a single request and response into durable, multi-step workflows. For anyone building on MCP, that is the more important change. Until now, a tool call was mostly one question and one answer. Support for longer-running, interruptible work is what lets an agent handle jobs that take minutes or hours.

Background from how the MCP spec is generally described, not from today's reporting: an elicitation lets a tool pause mid-job and ask you for input. A task lets a long operation be started, checked on, and collected later.

Status: REPORTED (shipped, per Yahoo and Forkast). STILL OPEN: exactly what the 'narrow' launch includes, and Anthropic's own release notes were not in the coverage. Cost: price not stated. Worth it for: builders already running MCP tools. Limit: a tool that interrupts you only helps if it asks at the right moments, and that design is on you.

What this means for you:

• Small business owner: a weekly job, like checking supplier stock, could pause and ask before it spends money.
• Job seeker or student: a small MCP workflow with checkpoints is a portfolio piece showing you can design for agents.
• Creator: a long batch job, such as captions for a set of videos, can check in with you instead of running unsupervised.

You're listening to SI Agent News. Still ahead: a Russian travel business making itself visible to AI agents, a shared memory project, and a security prompt.

Anthropic's platform team goes fully to Claude Code

TechGig reports that Anthropic's platform team has shifted entirely to AI coding with Claude Code. This is a dogfooding signal: a company using its own product for its own daily engineering. It carries real stakes for the engineers whose work changes, and it tells you the vendor is betting its own team's output on the tool. That is a stronger statement than a marketing page.

But the sourcing is thin, and 'entirely' carries a lot of weight. The report, as given, does not say whether it means no hand-written code at all, whether human code review is still part of the process, or how quality is measured. Status: REPORTED, single source. STILL OPEN: all three of those questions. Practical angle: if you run a team or a side project, do not copy the headline. Copy the discipline: decide what the AI writes, what a person reviews, and what you will measure. Before adopting anyone's setup, ask for numbers such as defect rates and review time (see the research section).

OpenAI reportedly detects widespread PHP vulnerabilities

The420.in reports that OpenAI detected widespread PHP vulnerabilities as worry grows about AI-driven cyber threats. PHP, a programming language, touches a huge share of the web per the editor's note. That is why this matters beyond developers. Many small business sites, online shops and blogs run on PHP-based software, often with nobody actively maintaining it. If AI tools can find weaknesses at scale, both defenders and attackers benefit, and slow-to-update sites are the easiest targets.

Status: REPORTED, one outlet. STILL OPEN: the scale, the severity, and whether fixes have landed. Practical angle: for any AI coding tool, Claude included, use it as a second pair of eyes on your own code, then verify the findings yourself. If you run a PHP-based site, check that your software and plugins are up to date and ask your host or developer what they are doing about it. The prompt later in this issue does exactly that kind of review.

A developer gave a TV a dose of Claude

Ifanr, a technology outlet, ran a piece titled 'A developer gave a TV a dose of Claude, and this is what changed in its system.' From the headline alone, a developer connected Claude to a television and the piece describes how the TV's system behaved differently afterward. The details given to us do not say what was changed, which TV or software was involved, or how Claude was connected, so we are not guessing at them.

Status: REPORTED, headline-level detail only. STILL OPEN: the mechanism and whether it can be repeated. Why it matters: it is a sign of AI reaching into everyday devices, not just chat windows and code editors. The useful takeaway for a hobbyist or tinkerer is to read the full piece for the method, and to think about permissions before you give any AI control over a device in your home.

Who is Dario Amodei?

The Profile, by Polina Pompliano, published a profile of Dario Amodei, the Anthropic CEO behind Claude. It is a background read, not news about a product or a change, and the details of the profile were not part of the material we were given, so we will not summarize claims it may make.

Status: REPORTED, profile piece. Why it matters: if you build on Claude or rely on it for work, knowing who runs the company and how they think helps you read its product and policy choices. It is lower priority than the tool stories above. Save it for a weekend, and treat it as context rather than guidance on what to do this week.

OneMemory: self-hosted shared memory for agents over MCP

A project on GitHub called OneMemory is described as self-hosted shared memory for AI agents over MCP. In plain terms, it is a memory store you run yourself that multiple agents can read from and write to through the same open standard covered above. The idea is that what one agent learns, another can use, without every tool keeping its own separate notes.

Status: REPORTED, an open-source listing. STILL OPEN: maturity, security model, and how well it works in practice, none of which the listing details give us. Cost: not stated; self-hosting means your own server time. Worth it for: builders who juggle several agents and keep re-explaining the same context. Limit: shared memory means shared mistakes, and anything stored there should be treated as sensitive. Test it with throwaway data first.

A small Russian travel agency connects to AI agents via MCP

A dev.to write-up describes connecting a small travel agency in a Russian region to AI agents via MCP, with 22 tools. The practical story is that a very small business exposed its own operations as tools an AI agent can call, rather than waiting for a big platform to do it. This is the 'make your business visible to agents' idea in a real, small-scale form.

Status: REPORTED, a developer's own account. STILL OPEN: results, costs and what the 22 tools actually do were not in the summary we have. Why it matters: for a small business owner, it shows the starting point is modest. List the few things a customer or assistant asks you most, such as availability, prices and bookings, and consider exposing those first. For a developer or student, it is a useful template for a real client project. Read the full post for the tool design before borrowing it.