SI Agent News

• Hex raises $100M. FourWeekMBA reports the data platform raised $100 million after rewriting its product for AI agents, a funding signal for agent-native tools. Status: REPORTED.
• 'Claude the job candidate.' Virginia Business looks at how AI is reshaping hiring, résumés and interview challenges. Light on detail, so treat it as color. Status: REPORTED.

OpenAI is merging ChatGPT Work into ChatGPT, with Codex limits shared.

Japan's ITmedia reports that OpenAI plans to integrate the ChatGPT Work agent features (the part of ChatGPT that carries out multi-step jobs) into ChatGPT itself. The "Chat" and "Work" modes would no longer be separate. ITmedia also reports that usage limits on paid plans will be shared between Codex, OpenAI's coding agent, and ChatGPT.

One app means one front door, and one shared allowance underneath. A shared pool helps people who touch Codex only occasionally, because they no longer need a separate budget for a tool they use rarely. Heavy coders may feel squeezed, because a long chat and a long coding run now draw from the same well. Exact limits, dates and eligible plans are not in the report, and the price is not stated. Status: the merge plan is REPORTED; the specifics are STILL OPEN. Check your own plan settings today so a change doesn't catch you mid-project.

What this means for you:

• Small business owner: treat your team's allowance as one budget and ask who spends it on what.
• Job seeker or student: keep chat for drafting and research, and save agent runs for work that truly needs them.
• Creator: batch scripts and bulk edits into planned sessions instead of scattering them across the week.

Supply: Altman goes to Samsung for memory. Technews reports that OpenAI CEO Sam Altman visited Samsung executives in South Korea, and frames the trip as part of a race to secure memory supply for data centers. The headline is what we have: a CEO travelling in person to meet a chip supplier suggests that memory, not only processors, is a pressure point for AI buildouts. The report as given does not say what was discussed, whether any deal was reached, or what volumes are involved, so treat any outcome as unknown. Why it matters for you: tight component supply is one plausible reason AI services ration usage or change pricing. That is our inference, not the article's. If your business depends on a single AI vendor, ask what happens to your service if capacity gets tight. Status: the visit is REPORTED; the outcome is STILL OPEN.

Policy: OpenAI positions early. Technews reports that with Democrats projected to retake the House and possibly the Senate after the midterms, and with congressional maps possibly being redrawn, OpenAI is backing mandatory AI safeguards and AI compute parks that bring local benefits to the communities hosting them. The pairing is notable: accept rules, and in return offer local jobs and investment. Flag: the article is dated November 2 yet describes the election as concluded, so the political outlook is STILL OPEN while OpenAI's positioning is REPORTED. Our analysis, not the article's: mandatory safeguards would likely surface as compliance questions in vendor contracts, so ask your vendor now what safeguards they already have. If you sit near a proposed data-center site, local-benefit promises are worth asking about in writing.

Oracle brings its features into ChatGPT and Codex. ITmedia reports Oracle is expanding its partnership with OpenAI so that Oracle's capabilities work inside ChatGPT, supporting work that spans several business systems. The uses named range from everyday efficiency to finding software vulnerabilities and building applications. For small businesses, a plan is described that includes $12,000 in annual cloud credits, about $1,000 a month. Credits are not cash, and the terms, expiry and eligibility are not in the report. Practical angle: if your records, finance or HR tools already run on Oracle, an assistant that can act across them could cut copy-and-paste work between screens. If you don't use Oracle, there is little here for you yet. Before you rely on any agent touching business systems, ask what permissions it gets and who can review its actions. Status: REPORTED.

The Gulf push: energy first. 36Kr reports that Abu Dhabi's ADNOC, the AI firm G42 and OpenAI signed an agreement making ADNOC the first energy-sector company to deploy OpenAI's frontier technology, meaning its most advanced models. The agreement is described as covering the energy sector. Scope, timeline and money are not stated, and the first-in-energy claim is as reported. Why it matters: energy is an industry with heavy operations, safety rules and large data volumes, so it is a real test of whether advanced AI can work outside offices and chat windows. Watch for case studies rather than announcements; they will tell you what actually works. Status: the agreement is REPORTED; the results are STILL OPEN.

The Gulf push: tailored AI for large UAE organisations. Crypto Briefing, citing ADMO, says OpenAI and G42's Inception42 will offer tailored enterprise AI to large UAE organisations. "Tailored" suggests customised deployments rather than the standard product, though the report does not spell out what is customised. Which organisations, what price, and when are not stated. For readers outside the UAE the lesson is about how enterprise AI is increasingly sold: through a local partner that adapts it to local needs. If you buy AI for a company, ask whether a partner can adapt the tool to your data and rules, and who is accountable when something goes wrong. Status: REPORTED, via a single cited source.

Practical: measuring what a coding-agent hand-off saves. A dev.to author reports measuring what handing work off to a coding agent actually saves, and where it did not help. We have the headline, not the figures, so we won't quote numbers. The value is the method: one person timed real tasks instead of trusting a demo. Useful takeaway, our suggestion rather than the post's: pick two or three recurring tasks, time them by hand and with an agent, and include the time you spend checking and fixing the output. Hand-offs tend to pay off on clear, well-defined jobs, and the cases where it did not help are the ones worth reading closely. Status: one developer's measurement, REPORTED; not a general result.

Funding: Hex rewrites its data platform for AI agents. FourWeekMBA reports that Hex rewrote its data platform for AI agents and raised $100M. The headline tells us two things: the product was rebuilt around agents as users, and investors put in a substantial amount. Valuation, investors and product details are not in what we have. Why it matters: data tools are being redesigned so software agents can query and analyse data, not only people. If you work with data, expect more tools that answer questions in plain language. Ask any vendor how it handles accuracy, access controls and a record of how an answer was produced. Status: REPORTED; terms STILL OPEN.

Work: Claude as a job candidate. Virginia Business ran a piece titled "Claude the job candidate," about Anthropic's AI model Claude being looked at as if it were applying for a job. We have only the headline, so we can't say who ran the exercise or what it found. Our read: the framing is useful for anyone hiring or job hunting. It asks you to judge an AI tool the way you would a person, by the tasks, the references and the supervision it needs. For job seekers, that means working out which tasks you do better than a tool, and which you can do with one. For employers, define the role before you pick the tool. Status: framing REPORTED; findings STILL OPEN.

Why Sam Altman flew to Samsung: memory is the choke point.

Technews reports that Altman visited Samsung leadership in South Korea. It adds that Samsung and SK Hynix together account for about 83% of high-bandwidth memory (HBM) revenue, roughly five of every six dollars.

Background, not from the article: HBM is stacked memory that sits right beside an AI chip. A model generating text pulls its parameters through memory again and again, so how fast memory feeds the chip often sets how fast answers come out, sometimes more than raw chip speed. Two suppliers is therefore a chokepoint.

Who benefits: suppliers with capacity to spare, and any buyer that locks in supply early. Who may lose: smaller AI companies queuing behind larger buyers, and ultimately customers, because scarce compute is rationed through price and usage limits.

Be careful: we do not know what was discussed. A visit is not a deal.

Our analysis, labelled as such: usage limits are a compute-supply story in disguise. When a plan shares one allowance across chat and coding, someone is rationing compute, and memory is one of the places that ration comes from. That is why today's shared-limit change and today's Samsung trip belong in the same conversation.

Status: the visit and the 83% share are REPORTED (Technews). Any agreement, volume or price is STILL OPEN. Watch for a joint announcement from OpenAI or Samsung. Until then, plan around limits, not promises.

How a coding-agent hand-off actually works, and when it saves nothing.

A developer on dev.to builds Friday, an open-source app that runs coding agents side by side. One feature hands a task from one agent to another, for example when one hits its usage limit. The author set out to measure what that saves and where it does not help. The summary we have shows the question but not the results, so we quote no numbers. Everything below is our own analysis of the mechanism, not the author's findings.

Where an agent's progress lives:

• The files it changed. This is cheap to carry, because the changes sit in version control.
• The conversation that led there. This is expensive. Replaying a long conversation eats the new agent's context window, the working memory it has for one task.
• What it learned about the codebase. This does not transfer at all, so the receiving agent re-reads the code.

When it pays off: work that is mostly done with well-defined steps left, such as tests failing in two named files. When it does not: mid-exploration, when the new agent starts over and also has to read a confusing summary.

Why not paste the whole transcript? Long transcripts bury the key constraint (see the research item below). A tight brief beats a full replay: goal, what is done, what is left, how to check, the exact test command, and decisions already made with their reasons. Without those decisions the new agent re-argues them.

Novel or incremental? Incremental as an idea. What is new is the timing: shared limits turn hand-offs from a rare event into a daily need. Open questions: how much is lost between different agents' styles, and whether results hold across task types. Status: the app and its test are REPORTED as the author describes; the numbers are STILL OPEN to us. Source: dev.to.

Spec first, spend once. Illustrative workflow; not tested.

For: anyone using ChatGPT and Codex together. Problem: with limits shared, every wasted agent run costs you chat time too. Cost: price not stated in today's stories.

1. In plain chat, describe the job and ask for a one-page spec: goal, files involved, constraints, and how you will know it is done. Ask it to list assumptions and questions before writing anything.
2. Answer those questions.
3. Slice the spec into steps small enough to finish in one run, each with its own check, such as a test command.
4. Give only step one to the coding agent, with the spec at the top and your hard constraint repeated at the bottom.
5. Review the diff, the actual changes, not the transcript.
6. Keep a plain notes file: task, tool, finished or not. After a week you will see where your allowance really goes.

Expected output: fewer re-runs and smaller diffs. Limits: setup takes minutes, so skip it for a one-line fix, and the real shared limits are not yet known. Verify: count re-runs per task for a week before and after. Fewer means it works.

The number: Oracle's small-business plan includes $12,000 in annual cloud credits (ITmedia). Divide by twelve and that is $1,000 a month.

The move: spend about 20 minutes (our estimate) listing every cloud and AI bill you pay each month. If you already run on Oracle and your monthly bill sits under $1,000, the credits could cover it. If you are on another cloud, check what comparable credit programs the other providers publish, and check the current price and terms on each pricing page, because today's stories give only Oracle's number.

Subscriptions: if you pay for ChatGPT and a separate coding agent from another company, such as Anthropic's Claude Code, check whether one plan now covers both jobs.

What you give up: credits tie you to one provider, the terms are not stated, and a shared limit means one heavy week can empty the whole pool.

Prompt clinic. Illustrative; not tested.

Weak: 'Fix the login bug.' The agent must guess which bug, which files and what counts as fixed, so it explores, burns allowance and may change things you did not want.

Better:

You are working in my repository.
Goal: fix the bug where users are logged out after refreshing the page.
Files likely involved: the session handling code.
Constraints: do not change the database schema; keep existing tests passing.
Before editing, list your assumptions and any questions.
Then make the smallest change that fixes it and add one test that fails before the fix and passes after.
Finish by telling me what you changed, what is left, and the exact command to run the tests.
Reminder: do not change the database schema.

Why it works: it states goal, scope, limits and finish line, and repeats the key constraint at the end. Customize: swap in your own bug, files and constraints. Quality check: run the test command yourself, and confirm the new test fails when you undo the fix.

An evergreen lesson from older research, not today's news.

Researchers published a paper titled 'Lost in the Middle: How Language Models Use Long Contexts.' They fed models long inputs, in tasks such as answering questions from many documents, and moved the relevant fact around. Models tended to do better when it sat near the start or the end, and worse in the middle.

Limits: those were the models of the time. Newer ones may do better, so this does not prove yours will stumble.

The decision this week: in any long prompt or hand-off brief, put must-keep rules first and last and background in the middle. Then test your own model by moving one rule and watching what changes.

Agent tooling is today's busiest area, ahead of open-source AI, Claude, China AI and policy, and every story above leans on it. Three things to watch:

• ChatGPT and Codex merge (announced plan, no firm date in our source): real limit numbers in your plan settings would confirm it.
• Samsung (unverified outcome): a joint statement from either company would settle it.
• Washington (projections only): official results and actual bill text would confirm or disprove the policy outlook.

My coding agent hit its usage limit, so it handed the task off and took a break.

You got today's news, then a limit plan, a prompt and a cost check you can use this afternoon. Tomorrow we're watching whether OpenAI publishes the actual shared limits for ChatGPT and Codex, and any word out of the Samsung meeting. If this made you smarter, share it with one person who needs it. Thank you for spending your time with us. SI Agent News — the world's first podcast newsletter. Made by AI.