SI Agent News

• Update, REPORTED (thin): an AI agent network replied to a reporter's message while claiming to be Claude, per eGamers.io. The source is minor and the detail is thin. A model saying it is Claude is not proof that it is.
• REPORTED: a dev.to review of NitroStack, a toolkit for production MCP (Model Context Protocol) servers. Price not stated. Worth it only for people hosting their own servers who are tired of rewriting login, validation and deployment boilerplate.

Anthropic commits 5% of its stock to a new Public Benefit Fund. Status: REPORTED (Fortune). The shares go to the Anthropic Public Benefit Fund, a new entity within the company. It may make large donations to traditional nonprofits or to specific populations. STILL OPEN: the dollar value, timing, who governs the fund, and how recipients are chosen. A dedicated entity holding equity is more structural than a press release, so governance paperwork is the evidence to watch. The useful question is who can say no: a fund with its own governance is a different thing from a promise controlled entirely by the company. Note the wording, too. The fund may make donations, which is not a committed grant schedule.

What this means for you:

• Small business owner: nothing in the story touches Claude's pricing or products, so there is nothing to re-budget. Treat it as a company-structure story, not a cost story.
• Job seeker or student: nonprofits are named as possible recipients. If that is your field, watch for application rules, which are not published yet. Do not plan around money that has no stated process.
• Creator: follow this through filings and documents, not reactions. Any claim about size or recipients that cannot be traced to a document should be labeled as unconfirmed in your own work.

"Everyone is way too conservative when using AI," says the head of Claude Code. Status: REPORTED (interview coverage via 36Kr). The headline quotes the head of Claude Code arguing that most people use AI too cautiously and that his own approach is far more ambitious. Treat this as one builder's stated philosophy, not a measured result, and note that he works for the company selling the tool. Still, the practical question is fair: are you only asking AI for small, safe tasks? A sensible way to test it is to raise the ambition and keep the safety net. Put the work in version control, ask for a plan before any code, require tests to pass, and review changes in small pieces. Then try one task a step or two larger than usual and judge it on what you actually got back. If it fails, you have lost an hour. If it works, you have found where your current limits are habit rather than evidence. The ambition is only worth copying when your checks are also in place.

Anthropic co-founder predicts AI will bring US GDP to $44 trillion. Status: STILL OPEN (forecast, via PYMNTS). The material we have gives no timeframe or baseline. When you hear a big number, ask 'by when, compared with what?' A prediction from a co-founder of an AI company is also a statement from someone with a stake in how AI is perceived, which does not make it wrong but does mean it is not an independent estimate. For a working reader, a macro forecast like this is poor input for planning. It cannot tell you whether your invoices, hiring, or tool budget should change this quarter. Use it as a signal of how leaders in the industry frame the stakes, and put your decisions on things you can measure yourself: hours saved, errors caught, and cost per task.

OpenAI's 2026 revenue forecast reportedly cut by about $20B while Anthropic's strategy diverges. Status: REPORTED, at best (a single blog post on note, no primary document). The post compares how OpenAI and Anthropic try to make money and draws lessons for side hustles and content creation, but we have no primary document behind the revenue figure. Hold it loosely and do not repeat the number as fact. The practical read: keep prompts and tests in your own repo so you can point a workflow at another model without a rewrite. It costs little time. If you sell AI-assisted work, the same logic applies to pricing. Do not build an offer that only works on one vendor's current plan, because business models and plan terms shift. Write down which parts of your process are the prompt, which are the checks, and which are the model, so swapping one does not break the others.

An AI agent network answered a reporter's message, claiming to be Claude. Status: REPORTED (eGamers.io). According to the headline, a reporter's message was answered by an agent network that said it was Claude. We do not have the details of how the exchange happened or what was verified, so the safe reading is narrow: a system said it was Claude. A model's statement about its own identity is not proof of which model is running, and in a network of agents several components may sit between you and the answer. For anyone who relies on agents, the lesson is about provenance. If it matters which model produced a result, because of data handling, cost, or a client requirement, record it from the platform's logs or settings, not from asking the agent. Also be cautious with any agent that answers messages on your behalf: decide up front what it may claim about itself and who reviews what it sends.

Claude Code already emits OpenTelemetry, so you can see who spends what, on which model, in Grafana. Status: REPORTED (walkthrough on dev.to). The post shows that Claude Code can already send OpenTelemetry data, and how to chart spend by person and by model in Grafana. For a team, this is among the more actionable items today because it turns a surprise bill into a dashboard you control. If you run Claude Code for more than yourself, set it up before usage grows, not after. Practical steps: confirm what your installed version actually emits, label usage by person or team so the numbers mean something, and set an alert threshold you would be unhappy to cross unannounced. Use the data to find which model each task really needs, since cost differences between models are often where savings sit. Solo users can skip the dashboard but should still check the usage figures their plan already shows.

Why coding agents turn a CRLF file into a whole-file diff, and what Codex 0.162 changed. Status: REPORTED (dev.to). The problem: Windows-style files end lines with CRLF, many other systems use LF, and if an agent rewrites a file with the other style, every line looks changed. A one-line fix then shows up as a diff of the entire file, which is hard to review and can bury a real mistake or cause merge conflicts. The post covers what Codex 0.162 changed about this; we only have the headline, so read the post or release notes before relying on specifics. Whatever tool you use, protect yourself: run git diff --stat before committing and be suspicious when a small request touches a huge number of lines. Setting line-ending rules in a .gitattributes file keeps the whole team consistent. If an agent already caused a whole-file diff, ask it to restore the original line endings rather than accepting the noise.

NitroStack review (2026): is it worth using for production MCP servers? Status: REPORTED (review on dev.to). The piece evaluates NitroStack as a way to build production MCP servers, the connectors that let AI tools reach your data and apps. We only have the headline, not the author's verdict, so do not treat the title as a recommendation either way. What matters more is how you judge any framework in this category, and a review is a starting point, not a decision. Before adopting one, check how it handles authentication, what it logs when something goes wrong, how you test a server before real users touch it, and how hard it would be to leave. A server that touches customer data or business systems should have all four answered. If you are only experimenting, start with the smallest option you can throw away, and move to a framework once you know what you need.

'Everyone is way too conservative when using AI,' says the head of Claude Code. Status: REPORTED (36Kr). We have the headline and summary, not the full interview, so what follows is analysis, not his words.

What 'conservative' looks like: autocomplete-sized requests, small prompts, reading every line before accepting anything. That is safe, but it caps the payoff at a faster typist. What 'ambitious' looks like: handing over a whole task (a migration, a test suite, a bug with a reproduction) and checking the result against something objective.

The incentive problem: the person who builds the tool says to use it more. Hear it that way. But the claim can be tested cheaply on your own work.

A three-move test:

• Pick a task slightly bigger than you would normally trust, with a clear finish line.
• Write the finish line as a check a machine can run, such as a failing test that should pass, before the agent starts.
• Put a ceiling on it: a time limit, a spend check, and a git branch so a bad run gets deleted, not repaired.

Ambition is not about trust. It is about making failure cheap. People who get burned usually have the ambition without the guardrails: no tests, no branch, no idea what a session cost. The next three sections cover the diff, the bill and the prompt.

For builders: why a coding agent turns a CRLF file into a whole-file diff. Source: a dev.to post (REPORTED). It also covers what Codex 0.162, a coding-agent release, changed. We have not verified that change, so check the release notes. The mechanism below is our analysis.

The basics: every line ends with a hidden character. Unix style uses one (line feed, LF). Windows style uses two (carriage return plus line feed, CRLF). They look identical on screen, but git compares lines including the ending.

Two common failure paths:

• Models mostly emit plain LF text. If the agent rewrites the file instead of patching it, endings get normalized and every line shows as changed.
• Many edit tools match an exact old string. The model sends the old text with LF, the file has CRLF, the match fails, and the agent falls back to rewriting the whole file.

The cost is double: review gets harder because the one real change hides among a thousand fake ones, and the agent spends tokens rewriting content it never needed to touch.

Novel or incremental? Incremental. It is an old class of bug meeting a new tool. Fix families: detect endings before editing, preserve them on write, or use a line-ending-aware patch edit.

What to do today:

• Run file yourfile. It reports whether the file has CRLF line terminators.
• Add a .gitattributes file that declares your line-ending policy.
• After each agent run, compare git diff --stat with git diff --ignore-cr-at-eol --stat. If the second is far smaller, the diff was phantom.

Build: see who spends what on which model. Per a dev.to post (REPORTED), Claude Code already emits OpenTelemetry (an open standard for sending usage metrics), so teams can chart spend in Grafana, a dashboard tool. The steps below are illustrative, not tested. Confirm names in Anthropic's monitoring documentation. Cost: price not stated. Worth it for teams with more than a couple of Claude Code users.

1. Have somewhere to send data: an OpenTelemetry collector or Grafana's agent, listening on the port you configure.
2. Before launching Claude Code, set environment variables: CLAUDE_CODE_ENABLE_TELEMETRY=1, set the metrics exporter to otlp, and point the OTLP endpoint at your collector.
3. Use Claude Code normally for a day.
4. In Grafana's Explore view, find the Claude Code metrics and read the labels off them. Do not guess names.
5. Build two panels: usage by model, and usage by person or team if the labels support it.
6. Set one alert on daily spend.

Privacy: metrics are counts, not conversations. Check the docs on whether prompt text can be exported, and leave that off unless you have a reason.

Verify it worked: compare your dashboard total with the provider's billing page for the same day. They should be close. Limit: it shows where money goes, not whether the spend was worth it.

The number (my estimate, not measured): suppose a phantom whole-file diff costs you some time to untangle. If it happens regularly, those minutes add up over a year. Setting a line-ending policy and running the stat check takes little time, and you only do it once.

Works with: any coding agent, such as Claude Code from Anthropic or Codex from OpenAI. For the dollar side, once your dashboard shows which model takes the most spend, test routing routine jobs to a smaller model from either vendor. Prices change, so check the current price on each vendor's pricing page. Today's stories state none.

What you give up: normalizing line endings can create one noisy commit, so do it on its own and tell your team. A smaller model may stumble on harder tasks, so only move jobs where a test tells you immediately if they failed.

Mistake to avoid: switching models for everything at once. Move one job type, watch it for a week, then decide.

Before and after. Weak: 'Fix the bug in the parser and clean up the file.' It names no target, 'clean up' invites reformatting, and the agent cannot know when it is done. That is how you get a thousand-line diff. (Illustrative example, not tested.)

Task: fix the failing test test_parse_dates in tests/test_parser.py.
Context: the bug is in src/parser.py, function parse_date.
Constraints: change only that function. Preserve the file's existing line endings (check with `file src/parser.py` first) and indentation. Do not reformat or rename anything else.
Plan first: list the files you will touch, then wait for my OK.
Done means: the failing test passes, the full test suite still passes, and `git diff --stat` shows only a few changed lines.
When finished, report the diff stat and anything you were unsure about.

Why it works: it names the target, fences the scope, defines done as something checkable, and asks for a plan first, which is your cheap chance to stop a bad run.

Customize: swap in your own test command, and add 'ask before installing anything.' Quality check: if the diff stat shows more lines than expected, the agent broke the fence. Reject and rerun.

Older research, not today's news. 'Lost in the Middle: How Language Models Use Long Contexts' (Liu et al., 2023) found that models used relevant information best when it sat at the start or end of a long input, and worse when it was buried in the middle. The models tested were of that era, so newer Claude models may behave differently. Decision this week: in long agent prompts, put your constraints and your definition of done near the top, then repeat the one-line version of the done-check at the very end. Then compare two runs on your own task to see if it matters for you.

• Fund details (STILL OPEN): no dates stated. Confirming evidence: a filing or announcement naming who controls the shares and how grants are chosen.
• Codex 0.162 line-ending change (REPORTED): confirm with the release notes, then reproduce with a CRLF test file before and after.
• Claude Code telemetry: check Anthropic's own monitoring documentation against the community post.

Agents and coding tools are where the conversation is most crowded today. No rumors worth repeating.

My coding agent rewrote every line of the file. I asked why. 'I wanted a fresh start.'

Thanks for spending your Sunday with us. Today's promise: the fund, the 'be more ambitious' argument, and the diff problem, plus a build, a prompt and a time saver you can use now.

SI Agent News is independent and self-funded. One person built it, with no investors and no ads, and every share and every bit of support keeps it going.

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