SI Agent News

• A user on the OpenAI community forum requests an advanced fantasy character creator in ChatGPT. CONFIRMED: the request exists. STILL OPEN: any OpenAI plan to build it.
• A dev.to post describes Eidos, a tool aimed at spec drift and fragmented rules across coding agents including Codex and Claude Code. REPORTED by the author; price not stated.

OpenAI's funding target is reportedly cut, but the details are unclear. Shattered.io headlines the story 'OpenAI Cuts Funding Target $10B to $20B.' Ten billion to twenty billion would be an increase, so the headline is ambiguous, and the summary we have is one line. The page address also mentions a $1.25 trillion valuation, which nothing in the text confirms. Status: REPORTED by one outlet. Direction and size: STILL OPEN. Why it matters: funding news can touch pricing, usage limits and how fast products ship, though nothing here says it will. The practical move is not to guess at the number. Note which of your workflows depend on a single vendor, and check that you could switch within a week if terms changed. Wait for a second outlet before changing any plans.

What this means for you:

• Small business owner: keep a second AI provider tested as a backup.
• Job seeker or student: learn the workflow, not one vendor's logo.
• Creator: don't build your whole channel on one feature.

ChatGPT search reaches 185.6 million EU users. Crypto Briefing reports ChatGPT search averaged 185.6 million monthly active recipients in the European Union, up from 159.1 million. By our arithmetic that is about 26.5 million more, roughly one extra person for every six before. The hook is the Digital Services Act (the EU's online platform law), with OpenAI's obligations reported as near. Status: REPORTED. Practical effect on users: STILL OPEN. Why it matters: a figure this size means people are using AI answers the way they once used a search box. If you sell to European customers or publish content they read, some of them may meet your brand through a chatbot answer first. A quick test is to ask ChatGPT search the questions your customers would ask, and see whether your business appears and whether what it says is accurate.

Samsung executives meet Sam Altman. Asia Daily reports Samsung executives met the OpenAI CEO to drive company-wide AI transformation. No product, contract or price is stated. Status: meeting REPORTED; any deal STILL OPEN. When a giant standardizes its AI, suppliers and clients start expecting the same tools. A meeting is not an agreement, so treat it as a signal of direction and not a done deal. For a small team the useful lesson is about process: large companies tend to set a company-wide approach before buying tools one at a time. Ask your own team for your AI plan in one sentence. If nobody can give one, that gap is worth fixing before you add another subscription.

An AI agent's trade broke a limit, and the refusal told it what would fit. A dev.to post describes an agent whose trade went over a set limit. The refusal it received said what would fit, rather than only saying no. We have only the title and a short summary, so we can't confirm the setup or the results. Status: REPORTED by one developer post; details STILL OPEN. Why it matters: this is a design lesson you can use without trading anything. A bare 'denied' leaves an automated tool to guess or retry blindly. A refusal that states the boundary, such as the maximum allowed, lets the tool correct itself in one step. If you build automations that call other systems, write your error messages so they say what is allowed. Spending caps, send limits and approval thresholds all work this way.

Running website A/B tests from Claude Code with an MCP server. Another dev.to post covers running website A/B tests from Claude Code through an MCP server. MCP is a standard way for AI tools to connect to outside services. The summary we have does not name the testing platform or give any results. Status: tutorial REPORTED; fit for your stack STILL OPEN. Why it matters: A/B testing means showing two versions of a page and keeping the one that performs better, and it usually needs a dashboard and some setup. Controlling it from an AI coding tool could shorten the loop between an idea and a live test. If you already use Claude Code and run a site, read the post for the workflow. Start with one low-risk page, such as a headline or a button, and keep a person approving what goes live.

Eidos targets spec drift for AI coding agents. A dev.to piece, 'Managing Architectural Context and Spec Drift for AI Coding Agents with Eidos,' addresses a familiar problem. Spec drift is when code slowly stops matching what was originally agreed, and coding agents can make it worse because they only see what is in front of them. We have only the title and summary, so how Eidos works is not confirmed here. Status: tool and approach REPORTED; real-world results STILL OPEN. Why it matters: whatever tool you choose, the habit is portable. Write down your architecture decisions and rules in one place the agent reads at the start of each task. Review changes against that written spec, not against memory. Teams using agents for more than a weekend project will meet this problem sooner than they expect.

Forum request: an advanced fantasy character creator for ChatGPT. A post on community.openai.com asks for a more capable fantasy character creator inside ChatGPT. It is a user request, not an announced feature, and nothing suggests OpenAI has responded. Status: REQUEST only; any roadmap plan STILL OPEN. Why it matters: requests like this show what everyday users want from these tools, which is structured, repeatable creative output and not only chat. You don't have to wait for a product. Game masters, writers and hobbyist creators can build a reusable template now: a fixed list of fields such as name, background, strengths, flaws and goals, pasted at the start of each session. That gives consistent characters and makes your work easier to carry to another tool if a feature never arrives.

A trading agent hit its limit, and the refusal told it what would fit. A developer on dev.to gave an agent a trading key, a $25 limit, and a Bitcoin short that was too big. The agent was Codex CLI (OpenAI's command-line coding agent). The guardrail said no; the interesting part, per the post, is what the agent did next, because the refusal told it what size would fit.

Status: the post exists, CONFIRMED. The result is one developer's account with a cut-off summary, REPORTED. It is not a study.

Analysis, ours: a bare 'denied' is a dead end, and the agent retries, rephrases or quits. A refusal with a reason and an allowed alternative is feedback. A limit in a prompt is a request; a limit enforced by code outside the model is a rule.

Three things to copy:

• Name which rule fired.
• Show the limit and current usage.
• State the largest action that would pass.

Mistake to avoid: over-explaining. Don't reveal secrets, other users' data or how to bypass the rule. Give the safe size, not the loophole.

Cost and worth: Codex CLI price not stated. Worth studying for anyone letting an agent touch money, files or production systems.

For builders: how an agent runs A/B tests through an MCP server. A dev.to post describes running website A/B tests from Claude Code (Anthropic's coding agent) through an MCP server. The post's framing: agents can ship a landing page in minutes, but learning whether it converts better than the old one needs a split test on real traffic. Status: the post is CONFIRMED; the mechanism below is our general explanation, not the post's details.

MCP is the Model Context Protocol, an open standard for letting an AI agent call outside tools. An MCP server advertises a menu of tools and the inputs each needs. The agent reads the menu, picks a tool and fills in the inputs; the host application runs it and returns the result.

A split test: each visitor is randomly assigned to version A or B and stays on it. You count how many in each group do the thing you care about, like signing up, then use statistics to ask whether the gap is bigger than chance. The agent does not change that math. It cuts the clicking needed to set up and read the test.

Novel or incremental? Mostly incremental. A/B testing is old. What's new is closing the loop (build, launch, measure) in one session.

Limits: you need real traffic, and a small site may take weeks for a reliable difference. A human should pick the success metric and stopping rule before launch. Your agent must support MCP servers, so check its documentation. Price not stated. Worth it for people shipping pages often; not for very low traffic.

This is a lesson, not news. The workflow is illustrative and not tested by us. It is inspired by the Eidos post on dev.to, which describes fragmented rules across agents and spec drift. You don't need Eidos to get most of the benefit.

Who it is for: anyone using more than one coding agent, or one agent across many sessions.

1. Create one plain-text rules file at your project root. Check your tool's documentation for the file name it reads automatically; names differ.
2. Keep it to five to ten rules: stack, how to run tests, what never to touch, naming, definition of done.
3. Point every agent at it, either by the name each tool expects or a one-line pointer: 'read this file first.'
4. Add a decisions list: three lines for each design choice already made, e.g. 'we use database X, not Y.'
5. After each session, ask the agent which rule it followed, and update the file when reality changes.

Stack: [language, framework]
Run tests with: [command]
Never edit: [folders]
Decisions: [choice one]; [choice two]
Done means: tests pass, no new warnings

How to verify: give two different agents the same small task. If both follow your naming and test rules, it works; if not, the rule is too vague.

Limits: a long file dilutes the agent's attention, so prune monthly. Cost: your time; tool prices not stated.

Estimate, ours, not from any source: if you retype the same project rules at the start of each coding session, call it 3 minutes, 5 sessions a week.

A shared rules file replaces the retyping. It works with Codex CLI (from OpenAI), Claude Code (from Anthropic), or another agent that reads a project file. Prices: check the current price on each tool's pricing page; today's stories don't state them.

Check your own number: time your next three session setups and multiply. If it is under a minute, skip this.

What you give up: upkeep. A stale rules file is worse than none, because agents follow it confidently. Spend five minutes a month pruning it.

Prompt clinic (illustrative, not tested). Weak: 'Write me an A/B test for my landing page.' It tends to return generic ideas, several changes at once and no way to read the result.

Better:

You are helping me plan a split test for my landing page. Context: [what the page sells, visitors per week, the action that counts as a conversion]. Goal: improve [one metric]. Propose ONE change only, with a hypothesis in the form 'If we change X, then Y will improve because Z.' Then give: the single success metric; a guardrail metric that must not get worse; how long to run based on my traffic; the stopping rule, decided in advance; and what result would make us keep the old page. If my traffic is too low for a reliable result, say so and suggest another way to learn. Do not declare a winner.

Why it works: one change shows what caused the result. The guardrail metric stops you winning signups while losing quality. The advance stopping rule stops peeking. The low-traffic line lets the model say the test isn't worth running.

Customize: swap in your metric and traffic. Quality check: ask the model to restate your numbers; if it invents traffic figures, discard the answer.

Lesson, not from today's stories: the peeking problem. This is long-established statistical background, not a new paper. If you check a split test daily and stop the moment one version looks like the winner, you declare false winners more often than the test's stated error rate promises. A false winner is calling a difference real when it is just chance. Statistically significant means a gap chance alone would rarely produce.

Decision this week: before launching any test, write down how long it runs and what counts as a win, and don't call it early. If your traffic can't reach a clear answer in that time, skip the test and talk to customers instead.

Agentic AI (software that takes actions on your behalf) is the busiest topic area today, with policy and security close behind. Today's stories fit: platform rules reaching ChatGPT in Europe, agents operating under limits, and tools connecting through MCP.

• OpenAI funding: one outlet so far; STILL OPEN. A company statement or major-outlet confirmation would settle it.
• Digital Services Act obligations for OpenAI: reported as near, no date in our sources. A regulator notice would confirm.
• Samsung and OpenAI: meeting reported; a named product or agreement would confirm any deal.

The guardrail told my agent no. Politely. With measurements. It's now the best-informed employee I have.

That's it for today. Tomorrow we're watching for any official word on OpenAI's funding target and the Digital Services Act timeline for ChatGPT search. Today's takeaways: a shared rules file, a refusal that teaches, and a better test prompt. If this made you smarter, share it with one person who needs it. Thank you for spending part of your day with us.

SI Agent News, the podcast newsletter. Made by AI.