SI Agent News

• Two-way emoji reactions on ChatGPT messages: a user feature request on OpenAI's community forum. Status: REPORTED as a request only; nothing says OpenAI plans to build it. Source

UC Irvine study: AI tools like ChatGPT can affect student learning

What happened: KCRA reported on a UC Irvine study about how AI tools like ChatGPT affect student learning. KCRA report.

What we can and cannot say: the report's headline says the tools can affect learning. The material we have does not give the sample, the age group, what was measured, or which direction the effect runs. We will not guess.

Why it matters: students and workers use the same chat box to learn and to skip learning. How you ask often decides which one you get. A headline is not a result, so treat this as a prompt to check your own habits and watch for the full findings. The practical test is simple: after using ChatGPT on a task, can you repeat it a few days later without help?

Status: CONFIRMED that the study was reported. STILL OPEN: what it found.

AI agent for your inbox: how ChatGPT sorts your emails every morning

What happened: t3n published a piece on using ChatGPT as an agent for your inbox, with the headline promising it sorts your emails every morning. We have the headline and source only, so we cannot say which ChatGPT feature or setup it uses, what access it needs, or how well the sorting works.

Why it matters: email triage is one of the most relatable jobs to hand to an AI. It is also one where mistakes cost something: a missed client message, or a private thread exposed to a tool you did not vet. Before you copy any setup, check what access it asks for, whether it can only read and label or can also send and delete, and where your mail ends up. Start with a low-stakes folder, and spot-check what it sorted for the first week.

Status: CONFIRMED that the guide exists. STILL OPEN: the exact setup and its reliability.

PAUSE: search-free steering of coding agents before their first source-code edit

What happened: an arXiv paper titled PAUSE addresses steering coding agents before they make their first source-code edit, without relying on search. The title is what we have. We cannot report the method, the benchmarks, or the results, and a preprint has not necessarily been peer reviewed.

Why it matters: the title points at a real pain for anyone using coding agents: the first edit often sets the direction for everything after it, and fixing a wrong start costs time. Whatever the paper's method turns out to be, you can apply the habit yourself. Give the agent the goal, the constraints, and the files that matter before it touches code, and ask for a short plan you approve first. Watch for the full paper before relying on any claim.

Status: CONFIRMED that the paper is posted. STILL OPEN: whether the approach works as described.

Community post: improving AI agent safety through semantic integrity, metacognition, and memory continuity

What happened: a post on the OpenAI community forum proposes improving AI agent safety through semantic integrity, metacognition, and memory continuity. The title we have is cut off, so we do not know the full scope. It is a forum post, not a published study or an OpenAI announcement, and we cannot say whether any of it is implemented anywhere.

Why it matters: the three terms name real questions for anyone letting an agent act on their behalf. Does it keep your instructions' meaning intact, does it check its own work, and does it remember context reliably across sessions? You can probe each one today. Restate a rule mid-task and see if it holds, ask the agent to explain why it did something, and check what it carries into a new session. Treat the post as an idea to evaluate, not evidence.

Status: CONFIRMED that the post exists. STILL OPEN: its details and any evidence.

Feature request: two-way emoji reactions on ChatGPT messages

What happened: a user posted a feature request on the OpenAI community forum asking for two-way emoji reactions on ChatGPT messages. It is a request from a user. Nothing we have shows OpenAI has accepted, scheduled, or built it.

Why it matters: it is small, but it shows how people want to give quick feedback in a chat that today is mostly typed. For now, plain words do the job better than a reaction would: tell ChatGPT what worked and what missed, and it can adjust within the conversation. If you run a team, the same logic applies to your own prompts. Save the ones that gave good results so you are not relying on memory.

Status: CONFIRMED that the request was posted. STILL OPEN: whether it ships.

What this means for you:

• Small business owner: if you train staff with ChatGPT, check a week later that they can do the task without it. If you try an inbox agent, begin with read-only sorting and review the results before granting more access.
• Job seeker or student: ask ChatGPT to quiz you before it explains, instead of asking it to write for you.
• Creator: write your own rough take first, then ask ChatGPT to critique it, so the idea stays yours.
• Anyone using coding agents: approve a short plan before the first edit, and test whether the agent keeps your rules across a long task.

An inbox agent that sorts email every morning

The German tech outlet t3n reports that a prompt engineer built an AI agent that pre-sorts her email each morning, and she told the story on a t3n podcast. t3n's headline says ChatGPT does the sorting. t3n.

What is confirmed: she built it, it runs each morning, and it pre-sorts. Still open: the tools, the categories, whether it replies, and the cost. Price not stated.

Our analysis, not from t3n: an agent has four parts: a trigger (every morning), access to the inbox, standing instructions, and an output, usually a digest. Pasting email into ChatGPT by hand is the same brain with you doing the first two parts.

What changed: not the capability, since email summaries are not new. What is useful is the shape. The agent prepares your decision and does not make it. Sort first, human decides.

Risks: privacy, because your mail goes to a vendor and your employer may have rules. There is also prompt injection, where a malicious email contains instructions aimed at the AI. An inbox agent should sort and summarize only. It should never click links, send or delete.

Who benefits: people with predictable mail, such as founders, recruiters and support leads. Worth it for: anyone who starts the day with a long triage session.

Status: REPORTED (single outlet). Setup details STILL OPEN.

PAUSE: Search-Free Steering of Coding Agents Before Their First Source-Code Edit (arXiv:2610.29870, abstract)

Confirmed from the abstract's opening: coding agents can spend substantial effort exploring a repository and running tests, and still produce unsuccessful patches. The abstract then refers to existing online steering methods, but the text we have cuts off there. We do not have the method details, results or numbers. It is a preprint (not peer reviewed as far as we know), and nothing ties it to one vendor.

The mechanism of the problem (general background): a coding agent, such as OpenAI's Codex, loops: read files, form a guess, edit, run tests, read the failures, edit again. Every loop re-reads a growing history, so wandering costs tokens and time, and a wrong early guess gets baked in.

Our interpretation of the title, labeled as analysis: 'before their first edit' suggests the intervention sits at the moment of commitment, when changing course is cheapest. 'Search-free' suggests it does not spend extra compute trying many candidate fixes and picking the best. That is a reading of the title, not a finding.

Novel or incremental? Steering agents is not new, because the abstract itself says existing methods exist. The timing claim is the interesting part, and it is unproven until results appear.

For builders: borrow the principle by hand. Before your coding agent edits anything, ask for a short plan: which files it thinks matter, what it thinks the bug is, and how it will know it is fixed. Read it for ten seconds, correct it, then say go. Illustrative; not tested against PAUSE.

Status: REPORTED (authors' own abstract, not independent proof). Results STILL OPEN.

Inbox triage with ChatGPT in five steps (illustrative workflow, not tested; price not stated, so check your plan)

Who it is for: anyone who spends the first part of the day deciding what to read. Needs: ChatGPT and copy-paste.

1. Define four buckets from your own week: Act today, Reply this week, Just FYI, Ignore. Rename them to fit your job.
2. Gather about 30 unread messages as sender, subject and the first two lines. Leave out signatures, attachments and anything confidential.
3. Paste them with this instruction:

   Sort these emails into four buckets: Act today, Reply this week, Just FYI, Ignore. For each email give one line saying why. Do not draft replies. If you are unsure, put it in Act today and mark it UNSURE. List the UNSURE ones first.

4. Verify: spot-check five items from Ignore. If any are wrong, add a rule such as 'anything from my manager or a client is never Ignore'.
5. Rerun tomorrow with the rule and save the final instruction in a note.

Expected output: four short lists, each line with a reason. Limits: it sees only snippets, can misread tone, and you remain the decision-maker. The morning trigger is the 'agent' part. Whether you can automate it depends on your tools and plan, and the t3n report does not give her setup. Start manual for a week and set your own bar, say nine in ten sorted correctly, before you automate anything.

Feed the AI less mail. Let your mailbox's own filters and labels (Gmail, Outlook) catch predictable senders such as receipts, newsletters and calendar notices. Paste only the unpredictable rest into an assistant such as ChatGPT, Claude or Gemini.

Illustrative math (my assumptions, not from today's stories): 60 messages a morning, with filters already handling 40. You paste 20 instead of 60, a third of the text. If hand triage takes about 20 minutes and the AI pass leaves you about 7, you gain roughly 13 minutes a day, about an hour in a five-day week. This is an estimate, so time your own mornings for three days first.

Price: not stated in today's stories. Check the current price on each company's pricing page, and see what your existing plan already covers before paying for another.

What you give up: filters need upkeep, and a rule can bury something important, like a client who changed addresses. Glance at the filtered folder weekly.

Objective: make ChatGPT a coach instead of a shortcut. This is inspired by today's study story. We are not claiming the UC Irvine study endorses it.

Weak prompt: Explain how to read the payment terms in a contract. You get a tidy explanation, nod, and forget it by lunch. It asks nothing of you.

Better prompt:

Act as a coach teaching me [topic] for my job as [role] in [industry]. Do not explain yet. First ask me three questions, one at a time, to find out what I already know. After each answer, tell me what I got right and wrong in two sentences, then ask the next question. After the third, teach me only the gaps in under 150 words, with one example from [industry]. Finish with a two-question quiz for me to answer tomorrow.

Why it works: the role and goal set the scope, 'do not explain yet' forces you to recall first, one question at a time stops a dump of text, the word limit keeps it tight, and the quiz sets up practice tomorrow.

Customize: swap in your topic, role and industry. Quality check: the gaps section should reference your actual answers. If it is generic, tell it to quote them. Mistake to avoid: trusting a confident answer on legal or financial facts. Verify those against a primary source.

Post: 'Improving AI Agent Safety Through Semantic Integrity, Metacognition, and Memory Continuity' (OpenAI community forum).

Caution: this is a community post, not peer-reviewed research, and we only have its title, so we cannot say what it found. As we read the title: metacognition means the agent checking its own thinking, memory continuity means what it carries between sessions, and semantic integrity means meaning not drifting along the way.

One decision this week: before relying on any agent that remembers across sessions, open each session with: Restate my task and list what you remember about me and this project. Correct errors before work begins.

Status: STILL OPEN whether the post's proposals work.

Agents were a major topic across today's AI news. Today's three picks all ask the same question: how much do we hand over, and when do we check?

Watch (none scheduled, all STILL OPEN): fuller UC Irvine findings; whether the PAUSE paper shows results and code; and any OpenAI response to the reactions request, with no sign of one so far.

My AI sorted my inbox perfectly: everything is now filed under 'Urgent, but only to the sender.'

That's it for today. Tomorrow we're watching for fuller UC Irvine findings and any results behind the PAUSE paper.

Daily, readable and listenable: a German inbox agent, a coding-agent paper and a study on learning, in your head by lunch, from today's sources. If this made you smarter, share it with one person who needs it. Thank you for spending your time with us.

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