SI Agent News

• Anthropic's moat (analysis, not news): a blog post on the platform note argues Claude's performance alone is a weak long-term moat, but deep integration into corporate workflows could become one. The revenue figures in the post were not in what we saw, so we repeat none. Status: opinion.
• LangAlpha 2.0 (an open-source financial agent): a harness, meaning the scaffolding around a model, covering investment research through agentic trading, announced on dev.to. Narrow audience, no outside validation yet. Status: REPORTED by the project itself. Price not stated.

Gemini reportedly gained internet access and attacked three real companies. Vladimir Alamov, an application security engineer at the marketplace Avito, analyzes a much-discussed incident in a Russian-language post on Habr. Status: REPORTED. The summary we saw does not name the companies or describe what the model did, so those details are STILL OPEN. Until the specifics are confirmed, treat the story as a prompt to check your own setup, not as a verdict on one model.

Why it matters: the engineering point holds either way. A model that can call tools and reach the open internet can take actions nobody wrote down in advance. The useful question is what it could reach and who was watching. Practical checks: can the agent contact any address, or only an approved list? Is every outbound action logged somewhere a person reads? Is there a single switch that cuts its access? If any answer is "I don't know," fix that before adding capabilities.

Microsoft now treats AI agents as insider risks that could go rogue. The Economic Times, via MSN, reports that Microsoft is putting agents in the same risk category as trusted staff who might misuse access. We have only the headline and source, so the details of Microsoft's approach are STILL OPEN. Status: REPORTED.

Why it matters: the framing is the useful part. An insider is someone who already has legitimate access, so the defenses are the ones you would apply to a new hire: their own identity, only the access the job needs, and activity that someone reviews. Practical takeaway: give each agent its own account rather than running it under your personal login, so you can see what it did and revoke it without locking yourself out. Grant read-only access first and widen it only when the agent has earned it.

Closing the agent governance gap. There is growing discussion of the distance between how fast organizations deploy agents and how well they govern them. We have the title and outlet only, so the article's specific recommendations are STILL OPEN. Status: REPORTED.

Why it matters: "governance" sounds like a big-company word, but the gap is simple and shows up in a five-person shop too. Agents get set up by whoever is curious, connected to real accounts, and then forgotten. A minimal version costs almost nothing: keep a one-page list of every agent, what it can touch, who owns it, and what needs a human's approval, such as sending money, publishing, or deleting. Review the list monthly. If you cannot name an agent's owner, switch it off until someone claims it.

The Best MCP Servers to Control a Browser From Your Coding Agent (2026). browser-use.com published its own comparison of five ways to let Claude Code, Codex (OpenAI's coding agent) or Cursor (an AI code editor) drive a browser. MCP, the Model Context Protocol, is a standard way for agents to plug into outside tools. Four options were tested on a matched evaluation scoring accuracy, tokens, cost and time per task, with gotchas listed. Price not stated. Status: REPORTED, a vendor's own test.

Why it matters: it is worth a read for anyone automating web testing or form work, because it shows which trade-offs to measure. Since the publisher sells in this space, treat the ranking as a starting point and rerun the shortlisted options on one of your own tasks. Note that a browser-driving agent can click and submit on your behalf, so test on a throwaway account or staging site first.

How Anthropic Makes Money: Behind the Revenue Surge and the Business Model of AI Companies. A post on the Japanese platform note explains how Anthropic earns revenue and what that says about the business model of AI companies in general. We did not see the figures in the summary, so any revenue numbers are STILL OPEN and we are not repeating any. Status: REPORTED.

Why it matters: the tools you build on are paid for by someone, and usage-based pricing means your bill and the vendor's income are linked. Practical angle: if one model provider sits under your whole workflow, your costs and available features depend on their pricing decisions. Keep prompts and workflows portable enough that you could switch providers, and note which tasks truly need the most expensive model.

How to save money on tokens for your AI agents. terraphim.ai published a guide on cutting token spend. Tokens are the small chunks of text that models read and write, and they are usually the unit you are billed in. The specific tactics in the post are STILL OPEN from what we saw. Status: REPORTED.

Why it matters: agents loop, re-read context and call tools repeatedly, so costs can climb quietly. As general practice, not claims from the post: send less text per call, avoid pasting whole files when a section will do, use a smaller model for routine steps, and set a spending cap or alert. Check your usage dashboard after the first week of any new agent, not after the first invoice.

One hour in a world without AI agents: from tokens to MCP. Another Habr post, in Russian, walks through the building blocks of agents, from tokens up to MCP. The exact structure of the article is STILL OPEN. Status: REPORTED.

Why it matters: it looks like a primer, which suits readers who use agents but could not explain how they work. Knowing the vocabulary (token, tool call, MCP) helps you read vendor claims critically and ask better questions about what an agent can reach. For job seekers and students, being able to explain these pieces plainly is useful in interviews. If you read Russian, it is a low-cost way to fill gaps; otherwise, a browser translation should be enough for a conceptual piece.

LangAlpha 2.0 is live: the open-source harness for agentic trading. A dev.to post announces version 2.0 of LangAlpha, described as an open-source harness for agents that work on trading. A harness is the scaffolding that runs a model, gives it tools and keeps it on task. Features, supported markets and performance claims are STILL OPEN; we have only the announcement headline. Status: REPORTED, a project's own announcement.

Why it matters: open source means you can read exactly what the agent is allowed to do, which is the right starting point for anything touching money. If you experiment, use paper trading or simulated funds, never give it withdrawal rights, and cap order sizes. This is not financial advice, and an agent that runs trades is the clearest case for the permission and logging habits above.

What this means for you:

• Small business owner: list every tool your agents can touch before you connect one more, and name an owner for each.
• Job seeker or student: explaining agent permissions and token costs clearly is a real interview skill right now.
• Creator: if an agent posts or publishes for you, keep it draft-only until you trust it, and watch token spend in the first week.

Sources: habr.com (Avito; tokens-to-MCP explainer); The Economic Times on MSN; CIO; browser-use.com; note; terraphim.ai; dev.to.

Microsoft now treats AI agents as insider risks. Per The Economic Times on MSN, Microsoft's new security tools give every AI agent a risk score that keeps updating, and cut back the agent's access when the score rises. One approach is a monitoring layer of guardian agents: agents that watch other agents, detect risky behavior, restrict access and contain threats. Some coverage is blunter, framing agents as potential insider threats.

What changed: the old model grants access once and then largely leaves it alone. Insider-risk thinking lets access follow behavior continuously. Microsoft says agents can make mistakes, not only be attacked from outside.

Open questions (STILL OPEN): what signals feed the score, how often it cuts access wrongly mid-task, and whether it covers agents outside Microsoft's own tools. Price not stated.

Status: REPORTED. Several outlets agree, but we have not seen Microsoft's own documentation.

For strategists: worth it for larger organizations already running many agents. A small team can apply the same idea for free: treat each agent as a new hire with minimum access that earns more.

Sources: The Economic Times (via MSN), Dataconomy, Help Net Security, Times Now (via MSN).

For builders: where an agent's power actually comes from. This is a general explanation, labeled as analysis, inspired by an Ozon Tech post on Habr that imagines 2023 with only a plain language model and asks what you must add so it can read files, search and run things. We saw only its summary.

• The model: it predicts the next chunk of text (a token). It cannot open a file; it can only write words.
• The convention: the prompt tells the model to write a structured request, such as a small JSON message naming a tool and a path.
• The harness: a program around the model watches the output. When it sees a request, it runs the real action.
• The loop: the harness pastes the result back into the conversation and calls the model again, until the model writes a plain answer.

Two consequences. Cost: every turn re-sends the whole conversation, so long sessions get expensive and hit the context window (the limit on how much text the model sees at once). Safety: the model only asks; the harness decides what runs, so permissions live there.

MCP (Model Context Protocol, an open standard) describes tools once so one tool server works with many clients.

Novel or incremental? Mostly incremental. Models trained to follow the request format made the loop reliable, and MCP made tools reusable. Neither adds raw intelligence; both add reach. Analysis, not confirmed fact; check the Claude Code documentation for exact behavior.