When goods report themselves
On May 28, Wiliot and AT&T said they are scaling a supply-chain system where tiny battery-free tags sit on cases, pallets, or products. The tags send location, temperature, and status into AI systems, so a warehouse no longer waits for someone to scan every barcode.
A shipment can arrive, match the order, flag a warm cold-chain item, or show that it is in the wrong zone. For retailers, grocers, and restaurants, receiving and freshness checks move from end-of-day cleanup to live exception work.
People still handle disputes, safety calls, and process changes. The practical lesson is simple. In physical work, better AI often starts with a better live signal from the real world.
On May 28, Wiliot and AT&T said they are scaling a supply-chain system where tiny battery-free tags sit on cases, pallets, or products. The tags send location, temperature, and status into AI systems, so a warehouse no longer waits for someone to scan every barcode.
A shipment can arrive, match the order, flag a warm cold-chain item, or show that it is in the wrong zone. For retailers, grocers, and restaurants, receiving and freshness checks move from end-of-day cleanup to live exception work.
People still handle disputes, safety calls, and process changes. The practical lesson is simple. In physical work, better AI often starts with a better live signal from the real world.
AI can now test giant cargo routes before a shipper books the move
A shipper moving a turbine or power transformer used to wait weeks for clearance studies and railroad calls.
At Port NOLA, a digital rail twin lets teams enter size, weight and railcar data, then see possible routes almost at once. Engineers still approve the move, but the risky question comes earlier in the deal.
A shipper moving a turbine or power transformer used to wait weeks for clearance studies and railroad calls.
At Port NOLA, a digital rail twin lets teams enter size, weight and railcar data, then see possible routes almost at once. Engineers still approve the move, but the risky question comes earlier in the deal.
Make AI rehearse a data deletion request before a real customer does
A company says, "You can delete your data." The awkward question is not whether the privacy policy sounds reasonable. It is where that promise actually lands.
This is a strong assignment for AI: take one fake or consented test user, then read the privacy policy, deletion workflow, data map, CRM and billing exports, support macros, product logs, vendor list, backup rules, and any AI agent memory or workflow traces. Ask it to simulate the request from intake to closure and return a blocker map: system checked, record found, evidence missing, owner, approval needed, and the sentence in the policy that created the obligation.
The useful output is not a prettier legal reply. It is a rehearsal report showing that the customer profile is easy to remove, but the support ticket keeps a copy, the analytics export has no owner, a billing note is exempt, a backup rule is vague, and the new support agent stores context nobody put in the data map.
That changes the mental model of AI. The assistant is not just answering a privacy question. With context and limited tool access, it can walk the organization and test whether a public promise survives contact with real systems.
The boundary is exactly why this works. AI should not delete production records, certify compliance, or send the customer response. Use read only access, minimized sample data, logs of the rehearsal, and human approval from privacy, legal, security, and system owners. The machine can find the cracks; accountable people still decide what must be removed, retained, or rewritten.
A company says, "You can delete your data." The awkward question is not whether the privacy policy sounds reasonable. It is where that promise actually lands.
This is a strong assignment for AI: take one fake or consented test user, then read the privacy policy, deletion workflow, data map, CRM and billing exports, support macros, product logs, vendor list, backup rules, and any AI agent memory or workflow traces. Ask it to simulate the request from intake to closure and return a blocker map: system checked, record found, evidence missing, owner, approval needed, and the sentence in the policy that created the obligation.
The useful output is not a prettier legal reply. It is a rehearsal report showing that the customer profile is easy to remove, but the support ticket keeps a copy, the analytics export has no owner, a billing note is exempt, a backup rule is vague, and the new support agent stores context nobody put in the data map.
That changes the mental model of AI. The assistant is not just answering a privacy question. With context and limited tool access, it can walk the organization and test whether a public promise survives contact with real systems.
The boundary is exactly why this works. AI should not delete production records, certify compliance, or send the customer response. Use read only access, minimized sample data, logs of the rehearsal, and human approval from privacy, legal, security, and system owners. The machine can find the cracks; accountable people still decide what must be removed, retained, or rewritten.
Physics AI moves into real engineering loops
Mistral has agreed to buy Emmi AI and is folding physics-AI models into industrial workflows: airflow, crash tests, chip equipment, control loops and digital twins. In its physics AI announcement, the company says the aim is to predict physical behavior from geometry, boundary conditions or sensor data in seconds, then use CFD and FEM solvers for verification and edge cases.
That changes the design rhythm. Instead of waiting hours or weeks to test a few variants, engineers can screen many shapes, materials or operating states before spending server time, solver licenses and expert review on the best candidates.
This affects aerospace, automotive, energy, semiconductor and industrial equipment teams first. The limit is sharp: fast prediction is not certification. Safety, production and regulation still need trusted solvers, real tests and accountable engineers.
Mistral has agreed to buy Emmi AI and is folding physics-AI models into industrial workflows: airflow, crash tests, chip equipment, control loops and digital twins. In its physics AI announcement, the company says the aim is to predict physical behavior from geometry, boundary conditions or sensor data in seconds, then use CFD and FEM solvers for verification and edge cases.
That changes the design rhythm. Instead of waiting hours or weeks to test a few variants, engineers can screen many shapes, materials or operating states before spending server time, solver licenses and expert review on the best candidates.
This affects aerospace, automotive, energy, semiconductor and industrial equipment teams first. The limit is sharp: fast prediction is not certification. Safety, production and regulation still need trusted solvers, real tests and accountable engineers.
Do not enter a vendor renewal call armed only with vibes
The useful AI assignment is not "summarize this contract." It is "build the renewal dossier." Give it the signed agreement, invoices, usage export, admin logs, SLA terms, support tickets, security questionnaire, a few competitor pages, and the Slack or email complaints people usually forget until the meeting starts.
The output should be a buyer's packet: what you paid for, what was actually used, which promises were missed, where support consumed time, what switching would break, what competitors now offer, and which negotiation asks are backed by evidence. It can also draft the procurement email, but that is the least interesting part.
This changes the shape of AI at work. The assistant is not just a chatbot with better prose. With enough context, it becomes the analyst who connects finance, legal, security, operations, and user pain into one reviewable artifact. A renewal stops being memory theatre and becomes an evidence review.
The boundary matters. AI should not invent leverage, interpret law as fact, terminate a vendor, or send the email without approval. Procurement, legal, finance, security, and the business owner still decide what is true, fair, confidential, and worth risking in the relationship.
The useful AI assignment is not "summarize this contract." It is "build the renewal dossier." Give it the signed agreement, invoices, usage export, admin logs, SLA terms, support tickets, security questionnaire, a few competitor pages, and the Slack or email complaints people usually forget until the meeting starts.
The output should be a buyer's packet: what you paid for, what was actually used, which promises were missed, where support consumed time, what switching would break, what competitors now offer, and which negotiation asks are backed by evidence. It can also draft the procurement email, but that is the least interesting part.
This changes the shape of AI at work. The assistant is not just a chatbot with better prose. With enough context, it becomes the analyst who connects finance, legal, security, operations, and user pain into one reviewable artifact. A renewal stops being memory theatre and becomes an evidence review.
The boundary matters. AI should not invent leverage, interpret law as fact, terminate a vendor, or send the email without approval. Procurement, legal, finance, security, and the business owner still decide what is true, fair, confidential, and worth risking in the relationship.
Ask AI to catch the missing lab details before a result goes stale
A promising result can die quietly in the notebook. The entry says "incubated as usual", the sample photo shows labels that do not match the spreadsheet, and the instrument export has a file name nobody can explain two weeks later.
That is a real job to hand to AI right after the run, while the bench memory is still warm. Give it the lab notes, protocol PDFs, instrument logs, reagent labels, sample photos, timestamps, spreadsheets, and result files. Ask it to reconstruct the exact protocol, then mark what is fact, what is inferred, and what is missing: temperature, lot number, sample ID, exposure time, calibration setting, filename, odd pause, suspicious mismatch.
The useful output is not a nicer methods section. It is a repeatability audit: a clean step sequence, a table of conflicts, missing parameter flags, source links, and questions a researcher should answer before the result becomes office folklore.
This changes the mental model of AI in science. It is less interesting as a fluent ghostwriter and more useful as a second lab memory, able to compare messy physical evidence with digital records and say, "Could another person actually repeat this?"
The boundary is strict. AI must not invent missing values, certify the finding, rewrite the official record in silence, or process sensitive patient or proprietary data in an unapproved system. A researcher, PI, QA reviewer, or lab manager still owns the final record and the scientific claim.
#AIAgents
A promising result can die quietly in the notebook. The entry says "incubated as usual", the sample photo shows labels that do not match the spreadsheet, and the instrument export has a file name nobody can explain two weeks later.
That is a real job to hand to AI right after the run, while the bench memory is still warm. Give it the lab notes, protocol PDFs, instrument logs, reagent labels, sample photos, timestamps, spreadsheets, and result files. Ask it to reconstruct the exact protocol, then mark what is fact, what is inferred, and what is missing: temperature, lot number, sample ID, exposure time, calibration setting, filename, odd pause, suspicious mismatch.
The useful output is not a nicer methods section. It is a repeatability audit: a clean step sequence, a table of conflicts, missing parameter flags, source links, and questions a researcher should answer before the result becomes office folklore.
This changes the mental model of AI in science. It is less interesting as a fluent ghostwriter and more useful as a second lab memory, able to compare messy physical evidence with digital records and say, "Could another person actually repeat this?"
The boundary is strict. AI must not invent missing values, certify the finding, rewrite the official record in silence, or process sensitive patient or proprietary data in an unapproved system. A researcher, PI, QA reviewer, or lab manager still owns the final record and the scientific claim.
#AIAgents
Home heart checks are becoming doctor-reviewed heart-failure warnings
Before, a cuff and watch made a long home diary. Coredio says its AI can turn those readings into pressure and flow signals for a clinician to review after discharge.
The FDA Breakthrough label can speed review, but it is not clearance. The AI does not change medicine alone.
Before, a cuff and watch made a long home diary. Coredio says its AI can turn those readings into pressure and flow signals for a clinician to review after discharge.
The FDA Breakthrough label can speed review, but it is not clearance. The AI does not change medicine alone.
Codex can now test the Windows UI it used to only describe
OpenAI has added Windows Computer Use to Codex, so Windows developers and QA teams can let the agent see, click and type inside apps for debugging and tests. The Codex Computer Use docs draw the line: use it when files, commands or APIs are not enough.
The shift is in QA. An installer that fails, a settings screen that hides a button, or a desktop bug that appears after three clicks no longer has to be narrated to the agent. Codex can inspect the same interface a user would touch, while a human steers and approves.
The catch: on Windows it uses the active desktop, can take over pointer and keyboard, and needs the machine awake. It is unavailable in the EEA, UK and Switzerland at launch; enterprise access is default-off. Secrets and sensitive apps still need human discipline.
OpenAI has added Windows Computer Use to Codex, so Windows developers and QA teams can let the agent see, click and type inside apps for debugging and tests. The Codex Computer Use docs draw the line: use it when files, commands or APIs are not enough.
The shift is in QA. An installer that fails, a settings screen that hides a button, or a desktop bug that appears after three clicks no longer has to be narrated to the agent. Codex can inspect the same interface a user would touch, while a human steers and approves.
The catch: on Windows it uses the active desktop, can take over pointer and keyboard, and needs the machine awake. It is unavailable in the EEA, UK and Switzerland at launch; enterprise access is default-off. Secrets and sensitive apps still need human discipline.
Copilot is moving inside the Office document
Microsoft is redesigning Microsoft 365 Copilot so the assistant is no longer just a chat pane beside Word, Excel, PowerPoint and Outlook. In its announcement, the company shows a single entry point across apps, task agents and Copilot actions inside a paragraph, cell or slide.
That changes the workflow more than another prompt box would. A worker can ask for an edit where the sentence lives, check a spreadsheet cell in context, or reshape a slide while the artifact stays on screen. Managers and IT teams should judge these tools by context, permissions, audit trails and reversible edits, not only by answer quality.
The boundary is review. In-document AI can make a wrong change look finished, especially in financial sheets, client decks or confidential files. Microsoft cites early usage gains after rollout, but short internal windows are not proof of lasting adoption.
Microsoft is redesigning Microsoft 365 Copilot so the assistant is no longer just a chat pane beside Word, Excel, PowerPoint and Outlook. In its announcement, the company shows a single entry point across apps, task agents and Copilot actions inside a paragraph, cell or slide.
That changes the workflow more than another prompt box would. A worker can ask for an edit where the sentence lives, check a spreadsheet cell in context, or reshape a slide while the artifact stays on screen. Managers and IT teams should judge these tools by context, permissions, audit trails and reversible edits, not only by answer quality.
The boundary is review. In-document AI can make a wrong change look finished, especially in financial sheets, client decks or confidential files. Microsoft cites early usage gains after rollout, but short internal windows are not proof of lasting adoption.
AI video is turning into a conversation
Gemini Omni demos show a new edit loop. Feed a clip, an image, sound, and a prompt, then keep asking for changes. New setting. Different camera. Motion on the beat. Same scene, less slot machine.
For creators, the skill shifts from one perfect prompt to directing the remix. Use your own footage, track what changed, and do not fake real people.
Gemini Omni demos show a new edit loop. Feed a clip, an image, sound, and a prompt, then keep asking for changes. New setting. Different camera. Motion on the beat. Same scene, less slot machine.
For creators, the skill shifts from one perfect prompt to directing the remix. Use your own footage, track what changed, and do not fake real people.
Before onboarding anyone, make AI excavate the project folder first
Before a handoff, give an assistant that can read files a bounded folder and paste:
You want a source cited map, not a folder summary.
Keep access read only, exclude sensitive folders unless the tool is approved, and verify decisions with the owner before acting.
#PromptEngineering
Before a handoff, give an assistant that can read files a bounded folder and paste:
You have read-only access to this folder. Do not edit, move, delete, rename, or create files.
Act as a project archaeologist, not a summarizer. Build an onboarding index for someone joining tomorrow.
Return:
1. Project snapshot.
2. Key files and why each matters.
3. Made decisions: evidence, source path, confidence.
4. Contradictions with both source paths.
5. Stale assumptions and why.
6. Missing owners, open questions, and risky gaps.
7. The 10-file reading order.
8. Message to the owner asking for missing context.
Rules: cite file paths for facts, separate facts from inferences, mark uncertainty, list skipped files, and do not quote secrets or personal data.
End with: What I would verify with a human before acting.
You want a source cited map, not a folder summary.
Keep access read only, exclude sensitive folders unless the tool is approved, and verify decisions with the owner before acting.
#PromptEngineering
AI agents are moving student paperwork before staff open the case
A campus bot used to answer FAQs. Element451 says its higher ed agents now read transcripts, check applications for fraud signals, schedule meetings, and send follow ups inside existing student systems.
The boundary matters. Staff still make admissions and advising calls. The agent keeps the queue moving so a waiting student does not lose a week to inbox silence.
A campus bot used to answer FAQs. Element451 says its higher ed agents now read transcripts, check applications for fraud signals, schedule meetings, and send follow ups inside existing student systems.
The boundary matters. Staff still make admissions and advising calls. The agent keeps the queue moving so a waiting student does not lose a week to inbox silence.
Use Deep Research like an analyst, not like Google
Most people ask Deep Research to "find good sources".
That is only search with more words.
A better job is: make it build a small research file you can use to decide.
Use this prompt:
This turns research from a pile of links into a decision file.
Good for market research, product ideas, school work, content planning, hiring, buying software, and checking claims before you share them.
The trick is simple: do not ask AI to "search". Ask it to show evidence, weak spots, and the decision it supports.
Most people ask Deep Research to "find good sources".
That is only search with more words.
A better job is: make it build a small research file you can use to decide.
Use this prompt:
Act like a research analyst.
Topic:
[write the question]
My goal:
[what decision I need to make]
Return:
1. Short answer.
2. Main claims, with evidence for each claim.
3. Source quality: strong, medium, or weak.
4. What sources disagree about.
5. What is still unknown.
6. What I should check before trusting the answer.
7. A simple decision memo I can send to a teammate.
Rules:
Use plain English.
Separate facts from guesses.
Do not hide uncertainty.
Do not give advice without evidence.
This turns research from a pile of links into a decision file.
Good for market research, product ideas, school work, content planning, hiring, buying software, and checking claims before you share them.
The trick is simple: do not ask AI to "search". Ask it to show evidence, weak spots, and the decision it supports.
❤1💯1
AI is turning referral faxes into review packets
At Mayo Clinic's Rochester cardiovascular team, about 8,500 new patients a year enter through referrals, faxes, MyChart, calls, and missing records.
Now an intake system treats a fax as a trigger. AI pulls the patient name, referral reason, and insurance details, drafts next steps, and summarizes the packet for staff. People still verify the fields and make the clinical calls.
At Mayo Clinic's Rochester cardiovascular team, about 8,500 new patients a year enter through referrals, faxes, MyChart, calls, and missing records.
Now an intake system treats a fax as a trigger. AI pulls the patient name, referral reason, and insurance details, drafts next steps, and summarizes the packet for staff. People still verify the fields and make the clinical calls.
Voice AI helped shoppers finish orders without typing
Meesho built Vaani for shoppers in smaller Indian cities, where typed search and filters can block a sale. People ask in Hindi or English, compare items, clear doubts, then confirm the order.
Company data says 1.5 million users tried it in the first month and buyers who used it converted 22% more, with fewer returns and cancellations. The hard part is now trust, consent, and exact confirmation before money moves.
Meesho built Vaani for shoppers in smaller Indian cities, where typed search and filters can block a sale. People ask in Hindi or English, compare items, clear doubts, then confirm the order.
Company data says 1.5 million users tried it in the first month and buyers who used it converted 22% more, with fewer returns and cancellations. The hard part is now trust, consent, and exact confirmation before money moves.
Ask AI to investigate the spreadsheet number everyone is about to trust
The scariest number in a company is not the one that looks wrong. It is the one already pasted into the board pack, with enough decimals to look official and just enough history that nobody wants to touch it.
A useful assignment for AI is narrow: trace this one metric back to its parents. Give it the workbook, linked CSV exports, source tabs, email attachments, folder versions, meeting notes, and any change log it is allowed to read. Ask for a lineage report: final cell, formula path, source files, hardcoded edits, version mismatches, assumptions found in notes, and questions a finance or operations owner must answer.
That output is different from "summarize the spreadsheet". It turns the sheet from a flat table into a chain of evidence. The assistant is doing the annoying archaeology first, so humans can spend their attention on the judgment: whether the override was valid, whether the forecast rule still applies, whether the KPI should be shown at all.
The boundary is simple and non negotiable: AI does not certify the number. Use approved tools, limit access to the necessary folders, keep sensitive data out of consumer systems, and make the accountable owner sign off. The best answer is not confidence. It is a sourced risk log before a bad number becomes a decision.
#AIAgents
The scariest number in a company is not the one that looks wrong. It is the one already pasted into the board pack, with enough decimals to look official and just enough history that nobody wants to touch it.
A useful assignment for AI is narrow: trace this one metric back to its parents. Give it the workbook, linked CSV exports, source tabs, email attachments, folder versions, meeting notes, and any change log it is allowed to read. Ask for a lineage report: final cell, formula path, source files, hardcoded edits, version mismatches, assumptions found in notes, and questions a finance or operations owner must answer.
That output is different from "summarize the spreadsheet". It turns the sheet from a flat table into a chain of evidence. The assistant is doing the annoying archaeology first, so humans can spend their attention on the judgment: whether the override was valid, whether the forecast rule still applies, whether the KPI should be shown at all.
The boundary is simple and non negotiable: AI does not certify the number. Use approved tools, limit access to the necessary folders, keep sensitive data out of consumer systems, and make the accountable owner sign off. The best answer is not confidence. It is a sourced risk log before a bad number becomes a decision.
#AIAgents
A solar battery can now follow a goal instead of a schedule
Solar owners used to set charge windows, backup reserve, EV timing, and tariff rules by hand. Sigenergy's new SigenAgent asks for an outcome like lower bills or more backup power, then reads weather, prices, grid status, and device data to choose the next battery move. Critical settings still wait for user approval.
Solar owners used to set charge windows, backup reserve, EV timing, and tariff rules by hand. Sigenergy's new SigenAgent asks for an outcome like lower bills or more backup power, then reads weather, prices, grid status, and device data to choose the next battery move. Critical settings still wait for user approval.
Robot dogs are becoming weekend builds
A Shanghai hackathon put Unitree Go2 robots on DimOS, with Python, perception, navigation, memory, and an LLM agent runtime.
The wild part is the loop. Write behavior, replay it, debug it, test it on hardware, then watch real floors and camera noise break your plan.
For young builders, robotics starts to feel less like a locked lab and more like a repo you can actually poke.
A Shanghai hackathon put Unitree Go2 robots on DimOS, with Python, perception, navigation, memory, and an LLM agent runtime.
The wild part is the loop. Write behavior, replay it, debug it, test it on hardware, then watch real floors and camera noise break your plan.
For young builders, robotics starts to feel less like a locked lab and more like a repo you can actually poke.
🔥1
AI turns existing forklifts into a live warehouse risk map
Before, managers saw a crash report after the fact. Now a stereo camera and visual AI on each truck track indoor position, flag pedestrian close calls, and replay a shift without GPS, beacons, or floor markers.
Slamcore says the retrofit is already in 30+ facilities. The hard part is the human layer. Safety teams still need review rules, privacy limits, and operator trust.
Before, managers saw a crash report after the fact. Now a stereo camera and visual AI on each truck track indoor position, flag pedestrian close calls, and replay a shift without GPS, beacons, or floor markers.
Slamcore says the retrofit is already in 30+ facilities. The hard part is the human layer. Safety teams still need review rules, privacy limits, and operator trust.
Let AI turn your router mess into a readable home network map
The mystery device on your Wi-Fi list is usually handled with guesswork: unplug things, ask the family, hope "ESP32" is not a camera you forgot. A multimodal assistant can do something more useful if you treat it like a junior network admin with read-only evidence.
Give it the router export or screenshots, connected-device list, Wi-Fi names, photos of the router, switch, cables, device labels, and the security settings you are willing to share. The assignment is not "secure my network." It is: draw the map, identify likely devices and owners, mark unknowns, label wired ports, note weak settings, and rank the first checks a human should make.
The useful output is an operations artifact: a simple topology diagram, an inventory with confidence levels, a list of suspicious or stale devices, and questions such as "is this old printer still in use?" or "who owns the phone that last connected on Tuesday?" That is a different mental model of AI. It is not just answering a networking question; it is stitching together messy machine data and physical evidence into something a household or small office can act on.
The boundary is strict. Keep it read-only, redact passwords and public IP details, and do not let the assistant change router passwords, firewall rules, port forwarding, parental controls, or firmware settings. AI can prepare the map and the risk list. The owner still approves every change.
#CyberSecurity
The mystery device on your Wi-Fi list is usually handled with guesswork: unplug things, ask the family, hope "ESP32" is not a camera you forgot. A multimodal assistant can do something more useful if you treat it like a junior network admin with read-only evidence.
Give it the router export or screenshots, connected-device list, Wi-Fi names, photos of the router, switch, cables, device labels, and the security settings you are willing to share. The assignment is not "secure my network." It is: draw the map, identify likely devices and owners, mark unknowns, label wired ports, note weak settings, and rank the first checks a human should make.
The useful output is an operations artifact: a simple topology diagram, an inventory with confidence levels, a list of suspicious or stale devices, and questions such as "is this old printer still in use?" or "who owns the phone that last connected on Tuesday?" That is a different mental model of AI. It is not just answering a networking question; it is stitching together messy machine data and physical evidence into something a household or small office can act on.
The boundary is strict. Keep it read-only, redact passwords and public IP details, and do not let the assistant change router passwords, firewall rules, port forwarding, parental controls, or firmware settings. AI can prepare the map and the risk list. The owner still approves every change.
#CyberSecurity
A home AI agent can find your smart TV on Wi Fi, connect to it, mute it, turn it on, turn it off, and explain every step
I tested this on a real home network. The agent found a Samsung 65 inch QLED TV at 192.168.0.250. It used the Samsung remote WebSocket API. Mute was sent, power off was verified as standby, and power on was verified as on.
Small note: my test worked on the first try because Codex had already been allowed on this TV before. On a new TV, the first connection may show a request on the TV screen. Confirm it with the remote, then run the agent again.
If something does not work, ask the agent: what did you find, which step failed, and what should I confirm on the TV?
Prompt to run the agent:
Prompt to create a reusable skill:
Example after the skill exists:
Real result from my test:
TV found: Samsung 65 inch QLED.
IP: 192.168.0.250.
Power on: verified as on.
Mute: command delivered, but Samsung does not expose mute state in this API.
Power off: verified as standby.
The useful lesson is simple: the agent should not just press buttons. It should discover, connect, act, verify, and tell you where it is not fully sure.
I tested this on a real home network. The agent found a Samsung 65 inch QLED TV at 192.168.0.250. It used the Samsung remote WebSocket API. Mute was sent, power off was verified as standby, and power on was verified as on.
Small note: my test worked on the first try because Codex had already been allowed on this TV before. On a new TV, the first connection may show a request on the TV screen. Confirm it with the remote, then run the agent again.
If something does not work, ask the agent: what did you find, which step failed, and what should I confirm on the TV?
Prompt to run the agent:
Find and connect to the TV in my home. Mute it. If something goes wrong, stop and give me a short report with what you tried and what failed.
Prompt to create a reusable skill:
Create a Codex skill called home-tv-control. It must discover TVs on my local network, identify the TV model and IP, connect to Samsung SmartTV through the remote WebSocket API when possible, support status, mute, power-on, and power-off commands, and always report the result. If mute cannot be read back from the TV API, say that the command was sent but the mute state was not confirmed.
Example after the skill exists:
Use the home-tv-control skill. Find my TV, show its status, mute it, then turn it off. Give me the IP, model, command results, and final power state.
Real result from my test:
TV found: Samsung 65 inch QLED.
IP: 192.168.0.250.
Power on: verified as on.
Mute: command delivered, but Samsung does not expose mute state in this API.
Power off: verified as standby.
The useful lesson is simple: the agent should not just press buttons. It should discover, connect, act, verify, and tell you where it is not fully sure.