SI Agent News

• MCP security auditor (REPORTED, dev.to): a developer open-sourced an auditor that scans MCP tool code using an AST (abstract syntax tree, the code's structure laid out as a tree), plus a crash firewall for multimodal logs and a zero-downtime model router. Single author, no traction yet. Source
• LLM_Tails (REPORTED, note): a free open-source rule set telling coding agents such as Claude Code and Cursor not to build without permission, not to hallucinate and not to forget. Free; worth testing. Source
• Audio for video agents (REPORTED, dev.to): a Thai-language guide lists tools that let a Claude agent hear real audio while editing video. It says it was written by AI under human review, so verify before you build on it. Source

Anthropic has made new Claude Enterprise cost and usage tools generally available. On October 28, Cryptopolitan (via MSN) reported that Anthropic launched three Claude Enterprise features: insight into how employees use Claude, the cost of that use, and a third capability that the available summary cuts off before explaining. Yahoo adds that these tools went out in beta and are now generally available on Enterprise plans. The news is the release status. Beta to generally available means a finance or IT lead can plan around these tools instead of treating them as experiments. Cryptopolitan frames the launch against lingering doubts about whether AI spending pays back, and cost visibility is the first thing a skeptical budget owner asks for. Status: REPORTED by two outlets; Anthropic's own announcement is not in today's set. Price: not stated. Worth it for finance and IT leads already on Enterprise. Open question: usage dashboards show activity, not value. Someone still has to connect the numbers to outcomes. Cryptopolitan on MSN · Yahoo

What this means for you:

• Small business owner: if your company is on Enterprise, look at the cost view before your next renewal.
• Job seeker or student: being able to say how you measured an AI tool's value is a real résumé line.
• Creator: track which of your own prompts and tools you actually use, then drop the rest.

Copilot metrics now see skills, agents and MCP servers. Visual Studio Magazine reports that GitHub Copilot usage metrics can now identify custom skills, agents and MCP servers. Until now, the question admins could answer was who is using Copilot. This update moves toward which of the custom pieces your team built or installed are actually being used. That matters because custom agents and MCP servers cost time to build, maintain and secure, and an unused one is pure overhead. With this data an admin can retire dead weight, double down on what people reach for, and justify the effort of building more. It is the same theme as the Anthropic story: measure usage before you argue about value. We only have the headline-level report, so check your own admin settings for how the data is exposed. Price: not stated. Worth it for admins running Copilot at scale. Status: REPORTED. Visual Studio Magazine

Turo and New Relic: AI writes more code, so reliability becomes the bottleneck. New Relic published a piece with Turo titled "Turn AI code volume into production reliability." We only have the title and source, not the details, so treat the specifics as unknown. The framing is worth your attention anyway. When AI assistants let a team ship more code faster, the limiting factor shifts from writing to running: spotting what breaks, finding out why, and fixing it before customers notice. Note that this is a vendor publication, so it is marketing as well as information. The practical takeaway for any small team using AI to code is to invest in monitoring and alerting at the same pace as your output. More code with the same visibility means more problems you hear about from users first. Status: REPORTED by the vendor; no figures in our summary. Price: not stated. Source: newrelic.com.

An open-source toolkit for hardening production AI agents. A dev.to post describes open-sourcing three pieces: an MCP AST security auditor, a multimodal JSONL crash firewall, and a zero-downtime LLM router. From the title alone, these target three common production headaches: unsafe MCP server code, malformed or multimodal input crashing an agent, and keeping service up when you switch between language models. We have not seen the code or any test results, so we cannot say how well they work. Still, the categories are the useful part. If you run an agent that touches real systems, ask whether you have reviewed the MCP servers it connects to, whether bad input can take it down, and whether a model outage would stop your service. Review any open-source tool before putting it near production. Status: REPORTED by the author; unverified. Price: open source per the title. Source: dev.to.

Serverless MCP for business operations. Another dev.to article covers an architecture and implementation for running Model Context Protocol (MCP) servers in a serverless setup for business operations. MCP is the standard way to give an AI assistant access to tools and data, and serverless means you pay for usage rather than keeping a server running. For a small business, that combination is attractive: connect an assistant to your order, inventory or support systems without managing infrastructure. The caution is that every connection you expose is also something that needs permissions and logging. Our summary has no code or cost figures, so read the article's architecture section with your own security needs in mind before copying it. Pair it with the auditing story above. Status: REPORTED as a tutorial; not independently tested. Price: not stated. Source: dev.to.

A practical guide to AI coding agents: context, workflows, verification. A dev.to guide breaks working with AI coding agents into three parts: the context you give the agent, the workflow you run it in, and how you verify its output. That ordering is sound advice even without the full text. Agents do better when they are told about your project, tasks are broken into reviewable steps, and nothing is accepted without a check such as tests, a diff review or a run. The verification piece is the one beginners skip and the one that prevents most embarrassing mistakes. If you are a student or new developer, this is a more useful read than most model-release news. We have only the title and summary, so judge the advice against your own experience. Status: REPORTED; a guide, not a news event. Price: free to read. Source: dev.to.

LLM_Tails: a free tool to curb needless code and design oversights. A post on note introduces LLM_Tails, described as a free open-source project meant to reduce unnecessary code and design oversights in AI-generated work. Anyone who has watched an assistant add three helper functions where one line would do will see the appeal. We have no benchmarks or user reports, so we cannot tell you whether it delivers. What we can say is that the problem is real: AI output often grows in size without growing in quality, and extra code is extra maintenance. If you try it, compare the result against what you would have accepted by hand, and keep the review step. Status: REPORTED by the author; early and unverified. Price: free, open source. Source: note.

Claude can edit video, but can it hear the audio? A dev.to post lists three tools that help an AI agent hear real sound, starting from the complaint that Claude edits video nicely but the audio sounds odd. The underlying issue is that an agent cannot judge audio it cannot perceive, so it edits blind to levels, timing and quality. The post's answer is to add tools that give the agent a way to check the sound. For creators using agents on video, this is a reminder that your own ears are still the final check. We do not have the names or details of the three tools in our summary. Status: REPORTED as a how-to; unverified. Price: not stated. Source: dev.to.

Cash flow becomes the yardstick. Cailian Press, a Chinese financial outlet, argues that cash flow is now the most important measure for AI giants. We only have a summary and no figures. In plain words, it means money coming in versus going out, rather than model scores or user counts. It is the investor-side version of enterprise buyers wanting cost visibility. For you, the relevance is indirect: companies under pressure to show cash generation may change prices, limits or free tiers, so avoid building a workflow that depends on a generous plan staying generous. Status: REPORTED as an argument, not a measurement. Cailian Press

Gemini 4 Carbon: a top score nobody can use yet. The Chinese outlet tmtpost reports that Google's Gemini 4 Carbon took first place in a global ranking but is not open to the public. The summary names no benchmark, scores or dates, and the headline asks whether it can save face for Google rather than claiming it does. A score nobody can call is not something you can buy or build on, so there is nothing to act on yet. Keep judging models on your own tasks, and revisit when access opens and independent results appear. Status: STILL OPEN. tmtpost

Turo and New Relic: giving agents live production context. A New Relic blog post, titled 'Turn AI code volume into production reliability', is a case study of the car-sharing marketplace Turo. The claim: using the New Relic MCP server (MCP, the Model Context Protocol, is a standard way to plug tools into AI agents), Turo gave its agents live production context, cut incident triage to about 60 seconds, and scaled releases safely. New Relic

Status: REPORTED. It is the vendor's own post, and no before-number is given, so we cannot say how much faster 60 seconds is.

Why it matters (analysis, not the post's wording): when AI writes more code, more code ships and more can break. An agent without production data guesses. An agent connected to live telemetry can look at the actual errors.

What to copy this week:

• Connect an observability tool to your agent read-only first, so it can look but not change anything.
• Make triage its first job: what broke, when, and what changed.
• Time it. Measure your own before and after, because the only number that matters is yours.

Who should care: any team shipping AI-written code with an on-call rotation. Risk: live data is sensitive data. Decide what the agent can see, log what it queries, and keep humans on the fix. Price: not stated.

How a serverless MCP server works. A dev.to explainer describes a serverless MCP server for business operations as a cloud-hosted server speaking JSON-RPC 2.0 and Server-Sent Events. The article is generic, so the background below is general knowledge of the protocol, flagged as our framing. dev.to

• JSON-RPC 2.0: a simple message format. The client sends a request with a method name, parameters and an ID. The server replies with a result or an error carrying the same ID. In MCP, the first two methods you meet are one to list tools and one to call a tool.
• Server-Sent Events: a one-way stream from server to client over ordinary web connections, so a long job can send progress instead of silence.
• Serverless: your code runs only when a request arrives, scales automatically, and you pay for use rather than idle servers. Check the current price on each provider's pricing page.

The mismatch: serverless functions are short-lived and stateless, while streaming and sessions like long connections. Design rule: keep each tool call short, keep the server stateless, and store session memory outside it. Cold starts, the slower first call after a quiet spell, are the other cost.

Security comes first: whoever hosts the server decides what an agent can trigger, so authentication and logging come before speed.

Novel or incremental? Mostly packaging. The protocol is unchanged; serverless is a deployment choice. Worth it for small teams who want tools online without managing servers; skip it for long-running jobs. Status: the article's performance and cost claims are STILL OPEN; we did not verify them.

Context pack plus verification gate for Claude Code. A dev.to guide on coding agents names two failures that dominate real work: under-context (the agent guesses the API and conventions) and weak verification. The workflow below targets both. It is illustrative, not tested by us. dev.to guide

1. Create a file called CLAUDE.md in your project's root folder. Claude Code reads it at the start of a session. Keep it under a page: build and test commands, the library versions to use, naming rules.
2. Ask for a plan before code: the files it will read and the change it intends. Wait for your yes.
3. Define done as a command, such as 'the test suite passes and the linter is clean', not 'looks right'.
4. Make it run those commands and paste the real output.
5. Read the diff (the list of changed lines) before accepting.

How to verify it worked: for a week, track how often the first attempt passes your checks and how many times you corrected an invented function or option. If both drop, keep it.

Limits: more upfront minutes per task, a stale CLAUDE.md misleads, and tests only catch what they cover. Cost: your Claude plan; price not stated in today's stories.

Run a 30-minute usage review every month. Anthropic (Claude Enterprise usage and cost insight) and GitHub (Copilot usage metrics) have both shipped admin visibility, per today's reports.

1. Open whichever dashboard your plan includes.
2. List what nobody uses: custom skills, agents or MCP servers with no calls.
3. Retire them, or cancel seats nobody touches.

The number (our estimate, not a sourced figure): pulling the same facts by hand from invoices and team surveys can be slow, and a dashboard could shorten that. Your numbers will vary. For prices on either tool, check the current price on their pricing pages.

What you give up: dashboards show activity, not outcomes, so also ask the team what actually saved them time. Per the reports, Claude's tools sit on Enterprise plans, so check what your tier includes.

Prompt clinic. Weak prompt: Fix the bug in my checkout code. It has no context, no limits and no way to check, so the agent guesses the cause and the library, then says it is fixed.

Stronger version (copy-ready, illustrative):

You are working in my repository. Task: the discount code is not applied when the cart has more than one item. Before changing anything, do three things. One, list the files you plan to read and why. Two, state the cause you suspect and what evidence would confirm it. Three, stop and wait for my approval. Do not guess library behavior; read the installed version's documentation or source. After I approve, make the smallest change that fixes it, add one test that fails before your change and passes after, run the test command from CLAUDE.md, and paste the real output. If anything is unclear, ask me instead of assuming.

Why it works: reading files first fixes under-context, the approval stop keeps you in control, the failing test and pasted output fix verification, and the last line gives it permission to ask.

Customize: swap in your bug, name your test command, add rules from your CLAUDE.md. Quality check: run the new test yourself. If it passes before the fix, it proves nothing.

Today's set has no research paper, so this is a caution, not a finding. The coding-agent guide claims two failures dominate: under-context (the agent lacked information) and weak verification (nothing checked its work). That is a practitioner's view, not a study.

Decision this week: run your own small study. For five working days, log every time an agent goes wrong and label it under-context, weak verification or something else. If one label wins, fix that first. It costs a note file.

Three threads: visibility (Claude Enterprise and Copilot now show what gets used), money (the cash-flow argument is the investor-side version of the same question) and production context (agents plugged into live systems through MCP).

What to watch: Anthropic's own announcement spelling out the third Claude Enterprise feature (STILL OPEN), and whether Gemini 4 Carbon opens to the public with a named benchmark and scores (UNVERIFIED until then).

My agent finally gave me a usage report: it says the biggest cost is me re-running it.

That's it for today. Tomorrow we're watching whether Anthropic spells out the third Claude Enterprise feature, and whether Gemini 4 Carbon reaches the public. If this made you smarter, share it with one person who needs it. Thank you for reading and listening. Made by AI.