**99% of people didn't realize that OpenAI has made yet another breakthrough in AI engineering
**
- Completely new (hello, Grokbot), life-changing launches that I'll discuss in this article: what they are, why they matter, and how to set them up for 100% effectiveness Dots is the most interesting place to start. It gives you a personal agent with its own cloud computer and the ability to delegate background work. Meet Dots.
DevDay also brought new models, shared documents, coding updates, plugins, and tools for building agents into your own products.
before the alpha - subscribe to my substack for more fresh alpha - https://substack.com/@0xcodila
1. Create Your Dot and Give It a Job
https://x.com/OpenAI/status/2104980481876070819
Open Dots in ChatGPT on desktop and follow the introduction. You can add connections during setup or return to them later.
At launch, personal access is rolling out to adults on Pro 100, 200, and 500
but anyway check the current access requirements before buying a plan for this feature!
If your account doesn’t have Dots yet, Steps 5-8 still give you useful places to start, depending on your plan.
The launch version of Dots runs on
GPT‑6 Astra
And here's the moment when Dots beats GrokBot Grok 4.6 is really weak, and the only way to improve it is by connecting GPT and Claude directly to it
https://x.com/0xCodila/status/2104634929518641487
Sol is a separate model choice for Work, Codex, and the API
Now give your Dot a responsibility. "Help me be productive" leaves almost everything undefined.
Start here:
Help me coordinate the launch of [product] on [date].Use the launch brief, checklist, and conversations I explicitly share. Identify changed requirements, blockers, missing owners, and decisions I need to make.For your first task, return a launch status report with source links. Ask about missing access. Draft suggestions; get my approval before sending messages or changing shared files.
Read that first report closely. Does it know the right launch date? Can you open its sources? Has it confused a suggestion with an agreed decision?
Fix those misunderstandings before giving it a recurring job.
Your first win is a correct brief. Everything else builds on that.
2. Connect the Apps and Computer It Needs
For messaging, open your Dot’s profile and select Add.
The channel guide covers supported options, including Slack and Microsoft Teams.
For our launch, connect the source documents and the communication tools the task needs. Then ask your Dot to identify what it can actually access.
Locate the launch brief and checklist I shared. Give me their links and latest relevant changes. Tell me which requested sources you cannot read.
For websites that require a login, open the cloud computer from Computers in the Dot’s profile. Use the browser handoff or private sign-in flow.
- Its browser has its own sessions. Being signed in on your personal laptop doesn’t automatically sign the Dot in.
Enter credentials through the sign-in interface. Complete any account verification, then return control so it can continue.
Cloud work can continue with your laptop closed. Work on your own computer needs that computer available and the app running. Choose the environment accordingly.
One easy mistake: connecting Slack doesn’t create a standing instruction to watch a channel. We’ll set that up explicitly in Step 4.
3. Delegate Outcomes and Keep Context
A launch update touches several sources. Give your Dot the complete deliverable, with a clear definition of finished.
It can delegate background tasks. Those tasks receive relevant instructions and context; you shouldn’t assume every worker sees your entire conversation.
Try this once your connections work:
Compare the current launch brief, checklist, and selected Slack discussion.Produce one update with four sections: what changed, what is blocked, who owns the next action, and what needs my
decision.Link each factual claim to its source. If sources conflict, show the conflict. Label proposed owners as suggestions unless someone has accepted the task.
That last sentence matters. A beautifully formatted list can quietly turn a guess into someone else’s responsibility.
Use Activity to inspect delegated work and its outputs. Open the actual deliverable and check whether it answers the assignment.
- Dots can retain useful context through memory and notes. Buuut That doesn’t mean it preserves every detail of every conversation.
Keep important project decisions in an accessible source document, and point to it in future tasks.
For coding work delegated to the cloud, set up a Codex Cloud environment first. Step 7 covers that prerequisite.
The useful output here is one update you can act on, with the evidence close enough to check.
4. Make It Recurring-and Keep Control
Once the one-off report is useful, ask for a schedule.
Every weekday at 09:00 Europe/Sofia, prepare the launch update using the sources we verified. Continue until [end date].Deliver it to [supported destination]. Include changes since the previous report and decisions waiting on me. Confirm the saved schedule, timezone, and destination.
Check the entry under Scheduled. Confirm that it was actually created and that its timing matches your request.
- Event-driven monitoring is separate. Ask what the connected service supports, then verify the event and response before relying on it.
For example: a requirement changes in the launch channel → your Dot prepares an updated brief. Simply connecting the channel doesn’t establish that routine.
Next, review Custom rules under Settings → Personalization → Permissions.
You can express which actions should proceed, require an explicit request, ask for approval, or be handed back to you.
My starting rule for this launch would be:
Research and prepare drafts within the access I’ve granted. Ask before sending messages, editing shared project records, spending money, or publishing anything.
These rules guide behavior; they don’t grant missing app permissions or guarantee that every action will be handled perfectly.
There are also three different places to stop work:
- Pause - Activity - Scheduled
Pausing the main agent doesn’t automatically cancel the other two. Check all three when you’re shutting a workflow down.
5. Put GPT‑6.1 Sol to Work
The model picker is the next place I’d go.
GPT‑6.1 Sol is available in Work and Codex on Plus, Pro, Business, Enterprise, and Edu. Workspace administrators may need to enable it
https://x.com/thsottiaux/status/2105007628460109953
Also in chat model picker
OpenAI describes Sol as offering near-Astra performance at a lower cost. I’d compare them on a familiar, demanding task before choosing a default.
Open the model selector below the composer, choose Sol, and start with the default reasoning setting.
Give it a job with an output you can judge:
Read this launch brief and the current landing page. Find claims that are unsupported, ambiguous, or inconsistent. Suggest exact replacements and explain what evidence each change needs.
For API users, the model ID is gpt-6.1-sol.
Here’s the standard token-price comparison:
That makes Sol’s standard input and output unit prices 80% lower. Your final task cost still depends on token use, tools, and pricing conditions.
Sol supports a 1.05-million-token context window. Requests above 272,000 input tokens have higher rates; check the model page before treating a huge context as cheap.
Now separate model choice from speed choice.
Compared with standard Astra, Astra Ultrafast generates tokens up to 8× faster in Codex.
https://x.com/sama/status/2104994601140711896
A task with browser waits, tools, or long reasoning won’t automatically finish eight times faster
The speed guide lists Pro 500 and eligible Enterprise/Edu access. Ultrafast also consumes usage faster.
The new Pro 500 tier costs $500/month. Buying extra credits on Pro 100 or 200 doesn’t unlock Ultrafast.
Finally, Sign in with ChatGPT can let eligible Plus/Pro users apply their plan’s usage in participating third-party apps.
Choose the ChatGPT sign-in option, then enable plan usage where offered. Login support and plan-usage support are different; usage shares your allowance, and app fees may still apply.
6. Bring Your Work into ChatGPT Space
**Super interesting part
Our launch now has reports, decisions, and drafts. Give them a place where the team can find the current version.*
https://x.com/thsottiaux/status/2104983716049379472
At launch, Space and Pages are available on Pro, Business, and Enterprise.
Open Space, choose New page, and create a launch page. Add the brief, relevant files, and source links.
Use ChatGPT alongside the page to draft or revise it. The Space guide covers creating pages, organizing spaces, and sharing access.
Turn these launch materials into a working page with: current scope, approved claims, open decisions, owners, and a dated change log. Preserve the source links.
This is where Pages becomes useful: the agreed version has a home you can return to and update.
For a team space, use All → New → Space, name it, and invite collaborators. Check access before adding sensitive material; space membership applies across its pages.
Collaborative Slides is another DevDay announcement, with availability planned for the coming weeks. Treat it as an upcoming part of this workflow.
For shared automation, Teams and Team Tasks provide scheduled or event-driven work using configured team connections and service accounts.
- Bring in the workspace admin for that setup. A team workflow needs access that survives one colleague signing out or leaving.
OpenAI also announced @ChatGPT in Slack and Microsoft Teams. Administrators configure the integration and its permitted tools and channels.
- Your personal Dot in Slack and a workspace’s shared @ChatGPT integration have different setup and access rules. Decide whose sources the workflow should use.
Then there’s the Meetings plugin: install it from Plugins, finish audio setup, and use Take notes for a meeting.
Tell participants and obtain consent before recording. Afterwards, review the summary and proposed actions before turning them into commitments.
At launch, Meetings is a macOS desktop beta for Pro and Business, with Enterprise in alpha. Calendar connections add conveniences such as reminders.
7. Build, Review, and Ship with Codex
Suppose the launch report finds a real problem: the signup page breaks on mobile.
https://x.com/OpenAIDevs/status/2104996045482778973
Give the coding task a reproducible target.
First, configure Codex Cloud: choose Work in → Cloud, create an environment, and connect the required GitHub repository
Let setup inspect the project and install dependencies. Review its report, resolve gaps, and publish the environment before starting the task.
Reproduce the signup failure described in [issue]. Identify the cause, make the smallest appropriate fix, and run the relevant checks. Return the diff, results, and remaining uncertainties.
The refreshed Codex CLI provides another entry point. Follow the installation guide, sign in, and open it in your project directory.
To start with Sol:
1codex --model gpt-6.1-sol
The update adds voice control and a /agents view for tracking delegated work.
- The CLI also includes model selection, permissions, and review controls. Choose the environment that gives the agent the repository and tools it actually needs.
Next, open Code Review in the desktop sidebar, connect your provider, and select a pull request.
The review guide lists GitHub support and GitLab preview. Automatic reviews can take a first pass in the cloud once configured.
Review findings alongside the changes and test evidence before merging.
For security work, install Codex Security Cloud, select New scan, and configure the repository and cloud environment.
Enable ongoing commit checks when useful, and inspect the evidence for each finding.
- Fix with Codex can prepare a patch; review that patch before creating a draft PR.
My rule here is simple: ask for the evidence needed to accept the change, then actually read it
8. Build Your Own Tools with Sites and Plugins
After a few launch updates, you’ll notice repeated steps: the same inputs, the same format, the same checks.
That’s a good candidate for a plugin.
Where available, mention @Plugin Creator and describe the workflow. The creation guide explains how to refine, test, and install it.
Create a Launch Update plugin. Inputs: a launch brief, current checklist, and dated changes. Output: an update with sources, blockers, owners, and decisions.Ask for missing sources. Distinguish confirmed facts from suggestions. Prepare drafts for review. Use this attached update as the formatting reference.
Test it with incomplete information and conflicting dates. A workflow that only works on a perfect example won’t save much time.
The DevDay plugin announcements also cover submission and discovery, plus Extensions for richer interfaces such as sidebar apps, conversation panels, and file editors.
Explore the official Extensions examples before deciding whether your plugin needs a custom interface.
Next, use Sites to build a launch dashboard. Describe its users, source data, and the actions each person should be able to take.
The new Sites with plugins capability can use connected tools and data. At launch, those sites are workspace-private and depend on workspace enablement.
Each visitor uses their own connected accounts and permissions. Sharing the site doesn’t hand everyone your connections.
Preview with real permitted data. Save a version while iterating; deploying creates a live URL, so make publication a deliberate step.
MCP Events adds another piece: supported servers can deliver events that start agent work, using subscriptions and webhooks.
For example, a new launch blocker could trigger a draft update. The event guide covers the required server support; an ordinary connector doesn’t automatically become an event source.
Finally, Shareable Profiles give you a place to showcase selected Sites.
Open your profile, choose what to show, and review sharing. Personal profiles start private; availability and workspace controls vary, with Enterprise support listed as coming soon.
9. Build Agents into Your Own Product (Jev competitor)
https://x.com/thsottiaux/status/2104986448269279399
For developers, the next question is how to offer this kind of workflow inside an application customers already use.
The Agents API launched on September 10. DevDay expands the story with computer use; these are separate milestones.
Start with the official quickstart. Create a project API key with the required permissions, install the SDK, and run the supplied sandbox example.
Keep the key outside the agent’s sandbox. Your first goal is to create a session, observe progress, and inspect a real result.
Then add browser-based computer use if your workflow needs it. The guide covers website access requests, sign-in, browser activity, and session cleanup.
For our launch, start with a narrow task: check the public signup flow and report the first broken step. Use a test account when authentication is necessary.
- Decisions API chooses from predefined answers using text or image context. Its focus is classification, routing, and similar decisions.
Example design: route an incoming launch issue to copy, engineering, or human_review. Agree on the labels and evaluation examples before wiring the decision into production.
It launched in limited preview, with broader access planned over the following days. Check access before building a dependency around it.
For AWS teams, Bedrock Managed Agents brings OpenAI’s agent harness and model inference into Amazon Bedrock.
Its execution, authentication, and supporting services differ from the OpenAI-hosted API. Use the AWS-specific setup, including IAM, for that deployment path.
The Private Intelligence announcements also need careful reading.
Private Safety Processing supports automated safety review without OpenAI retaining the covered prompts and responses. Encrypted safety records remain in customer-controlled storage under the documented retention setup.
Private Inference was announced for preview this fall. Treat that as a future availability milestone when planning a deployment.
Finally, OpenAI Marketplace lets eligible enterprises use part of their OpenAI commitment for approved partner software. Access goes through the enterprise interest process.
Those launch statuses are recorded in the official DevDay recap.
10. Put It to Work: Four Practical Setups
You don’t need to assemble all of this on day one.
Pick the workflow that creates something useful for you this week. These are starting designs; each depends on the access and connections described above.
A. The launch coordinator
Give Dots the launch brief, checklist, selected discussions, and a verified schedule. Keep the current plan in a Space page.
Prepare today’s launch update. Show changes, evidence, blockers, and decisions that need me. Draft any follow-up messages for approval.
Check: Can you follow every important claim back to a source? Did the report catch the latest agreed change?
B. The creator’s research desk
Use a scheduled Dot task to collect changes from a defined source list. Use Sol to turn the verified notes into a draft.
Review these official sources for updates since [date]. Separate released features from previews and announcements. Link every factual claim. Suggest three article angles.
Check: Open the sources, verify dates, and remove any sentence that implies you personally tested something you haven’t used.
C. The developer’s issue-to-PR workflow
Give Codex a reproducible issue and a configured environment. Review the resulting changes, then use Code Review and security tooling where appropriate.
Reproduce this issue, propose a fix, and run the relevant checks. Show the diff and actual results. Flag anything you could not verify.
Check: Does the original failure disappear? Are the tests relevant? Can a reviewer understand the change and its remaining risks?
D. The team’s follow-up desk
Use Meetings for notes, Space for the agreed record, and a Team Task for follow-up once shared access is configured.
Extract decisions, proposed actions, owners, and dates from these notes. Mark anything uncertain. Prepare the follow-up for review before sending it.
Check: Did each owner accept the action? Are tentative dates clearly marked? Can the team open the linked record?
The Shift
The area where Dot really has an edge over GrokBot is the model.
But will that really bring a lot of people on board with OpenAI?
So you define the outcome → agents carry the work forward → you make the decisions.
The real alpha is connecting these releases into a workflow with a clear responsibility, useful context, and a result you can verify:
- Give Dots an ongoing job and clear permission boundaries.
- Use Sol and Codex to research, build, and review.
- Keep shared work in Space. Turn repeated steps into plugins and scheduled tasks
Now you're designing how the work runs.
What starts it? Which sources should it use? What does finished look like? Which decisions come back to you?
You now have the setup, the prompts, and four practical workflows. Pick one recurring job. Measure the time saved and the corrections needed. Then expand.
The skill to develop is defining work clearly enough that an agent can carry it forward without you directing every click.
Bookmark this playbook - then give your Dot its first real job
