←  Back to overview
Article

OpenAI DevDay: 25 releases, put to work

Twelve for our clients, eight for how we work at Nimble.

25 announcements.

That is what I counted on OpenAI’s DevDay recap.

A list that long is hard to act on.

So we gave every update one moment. A real job, in a context a client would recognise.

Here is what they add up to.

Your next customer might choose you without ever visiting your website.

They start in an AI assistant. They describe what they need. And they meet your service at the exact moment it becomes useful.

So the roadmap question changes: which valuable thing does your business do that an agent should help someone get done?

For an enterprise, the opportunity is to make its expertise easier to work with, wherever the customer starts.

Below: twelve client moments and eight changes to how we work at Nimble.

Every visual is an illustrative Nimble concept with synthetic data. None is a live deployment or an endorsement by the brands shown.

The 25 DevDay updates at a glance

Hold your horses: this is a lot. Twenty-five updates, twenty moments. Nobody reads that in one go, so tap the update you care about and we’ll scroll you straight there.

  1. Plugin creation, submission and discovery · Securex
  2. Plugin extensions · Walmart, Cigna (MCP Apps)
  3. MCP Events for plugin automations · ENGIE
  4. Decisions API · team.blue (Combell)
  5. Sign in with ChatGPT · team.blue Website Builder
  6. Computer use in the Agents API · Attentia
  7. Private Intelligence · Medi-Market, Maria Middelares
  8. Amazon Bedrock Managed Agents for OpenAI · An energy retailer
  9. Ultrafast · Loop Sleep
  10. Teams and shared tasks · Nimble · Team tasks
  11. Codex in the cloud · Nimble · Codex Cloud
  12. A refreshed Codex CLI · Nimble · Codex Cloud
  13. Code review · Nimble · Codex Security
  14. Codex Security Cloud · Nimble · Codex Security
  15. Dots · Nimble · Your dot
  16. Sites with connected plugins · Nimble · Sites
  17. @ChatGPT in Slack and Microsoft Teams · Nimble · Slack
  18. GPT-6.1 Sol · An accountancy firm
  19. ChatGPT Space · Nimble · Space
  20. Dynamic pages · Also worth knowing
  21. Collaborative presentations · Also worth knowing
  22. Meetings plugin · Also worth knowing
  23. Shareable profiles · Also worth knowing
  24. A new Pro plan · Also worth knowing
  25. OpenAI Marketplace · Nimble · Marketplace

Workspace agents (the energy-retailer switches) come from OpenAI’s developer documentation rather than the recap.

Twelve client moments

Plugin discovery · Securex

“Simon Verhaegen starts on 2 November as a full stack engineer, four days a week. What do I need to arrange?”

Tine runs People. She never asks for a plugin.

ChatGPT suggests a few that can help. She installs Securex, approves its scopes, and Simon’s onboarding file opens beside the chat. A payroll specialist validates the numbers. No contract or Dimona leaves the conversation.

Discovery is a new acquisition channel. Someone starting a side business has the intent, but no social secretariat yet. It now sits next to paid visibility, with ChatGPT Ads live in 60+ countries. Both only pay off if the service moves people forward.

The work is in the metadata. Test the indirect questions people actually ask (OpenAI’s guide).

An indirect hiring question, a suggested plugin, and the employer still deciding.
An indirect hiring question, a suggested plugin, and the employer still deciding.

Plugin extensions · Walmart

“Plan five easy dinners for four under $120. We already have rice and olive oil.”

A recipe list is a start. A basket you can edit is better.

The dinners stay in the chat. The basket opens in a panel beside it.

Quick actions sit right on the subtotal: swap salmon for chicken, skip the rice you already have, cook for five. The total updates. Nobody retypes the request. The flow stops at basket review.

Don’t rebuild the webshop. Meal plan to basket is enough to learn from.

Five dinners. One editable basket.
Five dinners. One editable basket.

MCP Apps · Cigna Healthcare

“In Barcelona 6–8 October, need a GP. What’s covered?”

Three checks hide in that question. Eligibility. Network. Pre-authorisation.

An MCP App answers with its own interface, inline. Each check gets its own chip. Three nearby practices on a map. The plan clause behind every statement. Data tools check; a separate render tool draws the card (OpenAI UI guidance).

A yes on one check never implies a yes on the others. And a care navigator gets the context, so the member never repeats their story.

Coverage, nearby options and the source in one card. No diagnosis, no booking.
Coverage, nearby options and the source in one card. No diagnosis, no booking.

MCP Events · ENGIE

21:40. A nudge arrives: electricity is cheaper between 01:00 and 05:00 tonight.

With MCP Events, a connected app sends signed events into a chat the user subscribed to (docs). The household picks its nudges: a cheap-hour window, a quick win, a payment reminder, a new charging plan.

The nudge brings the price curve, the cost and the saving. One button adjusts the plan.

Proactive design is mostly restraint. One delivery per event ID. Nothing scheduled without approval. Silence when nothing useful changed.

Proactive nudges the customer chose, with the numbers and a decision.
Proactive nudges the customer chose, with the numbers and a decision.

Workspace agents · an energy retailer

SW-4107 has been stuck for two days. The meter identifier is missing.

The retailer’s backend starts a published ChatGPT workspace agent by API. Retries with the same idempotency key are safe. Every run opens a conversation the team can follow (docs).

The agent prepares and routes. It never changes a contract.

Every blocked switch gets a reason, an owner and a next step.
Every blocked switch gets a reason, an owner and a next step.

Decisions API · team.blue (Combell)

“Since the migration, my customers no longer receive order confirmations.”

Hosting? DNS? Email? The customer shouldn’t have to know.

The Decisions API picks one answer from a fixed set, so routing stays inspectable. In this Combell graph, deterministic steps fetch context. Decision nodes pick the topic, spot the missing sending domain and choose the next action. A policy gate blocks any settings change. Low confidence goes to a person.

Small decisions, reused across brands, without flattening each brand’s service.

One request. The right next step, with every decision visible.
One request. The right next step, with every decision visible.

Sign in with ChatGPT · team.blue Website Builder

A florist decides tonight is the night. She opens a webshop on bloemen.be.

No new password. Signing in shares her name, e-mail address and profile photo. Never her chats or memory. Letting ChatGPT manage the shop later is a separate yes.

It removes a drop-off at the exact moment someone decides to start.

Open a shop without a new password. Further access is asked separately.
Open a shop without a new password. Further access is asked separately.

Computer use · Attentia

Some systems have no API worth the name.

An agent pulls the approved context by API, then fills a legacy payroll form for a move to a four-day week. The expert sees old and new values side by side. An unclear salary basis pauses the run. The form stays “Not saved” until the expert approves.

Computer use is a bridge, not the architecture.

The agent prepares the change. The expert decides what is saved.
The agent prepares the change. The expert decides what is saved.

Private Intelligence · Medi-Market and Maria Middelares

A customer at the counter. Early sixties, always tired, on a blood thinner. She points at the curcumin promotion.

The MediLife assistant flags it before anyone suggests a supplement, with the sources one tap away. The pharmacist confirms before anything is shared.

Grounded advice at the counter, with the pharmacist in charge.
Grounded advice at the counter, with the pharmacist in charge.

At Maria Middelares, a pre-operative consultation is two lanes that never merge. Administrative readiness: referral, questionnaire, practical information. Clinical review: a medication list that may be outdated stays with the clinician.

Private Intelligence (Zero Data Retention, Private Safety Processing, Private Inference in preview) is a building block. It is not an approval for every sensitive workflow.

Administrative completion never marks clinical readiness.
Administrative completion never marks clinical readiness.

Bedrock Managed Agents · an energy retailer

Step 3 of a contract change fails. Do you rerun all five?

Bedrock Managed Agents runs OpenAI’s agent harness and inference inside Amazon Bedrock, with tools in AgentCore or your own compute (docs). Give the agent its own identity, scoped permissions and approval points. Retry only the step that failed.

An agent you can’t operate is a demo.

A useful agent also needs an identity, an owner and a recovery plan.
A useful agent also needs an identity, an owner and a recovery plan.

GPT-6.1 Sol · an accountancy firm

A year-end file. A €38,400 roofing invoice, booked as repairs.

Rules can check that the trial balance ties out. They can’t tell an investment from an expense.

GPT-6.1 Sol reviews what rules can’t, with the invoice, the ledger and last year’s file as evidence. It offers close to Astra-level reasoning for a fifth of Astra’s token price, so the deep review runs on every file instead of a sample. The accountant decides how it is booked, and signs.

Put the stronger model on the judgement step. Then run it on everything.

Review every file, not a sample. Files and counts are illustrative.
Review every file, not a sample. Files and counts are illustrative.

Ultrafast · Loop Sleep

23:41. Someone can’t quiet their mind. The coach is still thinking.

Every second of silence pulls them out of the moment.

Ultrafast generates up to 8× faster in Codex and up to 6× in the API. GPT-6.1 Sol Ultrafast is coming soon. Use it on the live reply only; tonight’s soundscape runs on the standard model.

Pay for speed where a person is waiting. We do the same in our vibe-workshops: prototype while the group is still talking.

The coaching turn that cannot wait. Timings are illustrative.
The coaching turn that cannot wait. Timings are illustrative.

Build capabilities that outlast the interface

Preparing a hire. Finding care. Filling a basket. Routing a request.

Each is a capability: clear inputs, a source of truth, a result someone can verify. Your website can use it. So can your app, an internal agent and, with the right controls, someone else’s assistant.

That is what I mean by building capabilities, not just pages. The interface can change without rebuilding the service underneath each time.

It also answers platform dependence. Be discoverable anywhere. Keep the capability, and the handoff into your own channels, yours.

Eight ways Nimble works differently now

The same updates change how a small senior team builds. Product judgement matters more, not less.

Codex in the cloud

One brief. Three bounded tasks running in parallel in a shared cloud environment. Followed in the new /agents view, steered by voice from a phone.

Each task ends in one diff to review. Merge and release stay with us.

One brief. Three tasks in parallel.
One brief. Three tasks in parallel.

Codex Security Cloud and code review

A nightly scan finds 14 issues. Deduplication leaves 3. Investigation reproduces 1: a document read that skips the tenant check.

Codex writes the patch and the regression test. An engineer accepts, or dismisses with a reason.

Fewer findings. Each one verified.

From scheduled scan to a fix someone accepts.
From scheduled scan to a fix someone accepts.

Team tasks

Every Monday at 09:00, our product team Orion gets its weekly client update drafted. It reads Linear, GitHub and Figma, flags a design review without a ticket, and never posts without approval.

The team edits the instructions together.

Recurring work the whole team can steer.
Recurring work the whole team can steer.

Your dot

Our support agent NSUP-V2 gets one goal: every request reaches a reviewed next action within a working day.

Context, duplicates and drafts it handles on its own. Replies need approval. Production changes never happen without a person.

An illustration of the operating model, not a claim that a Dots integration already runs.

Support that keeps moving between conversations.
Support that keeps moving between conversations.

Sites with connected plugins

An account cockpit, built as a Site in ChatGPT, with Notion, Calendar and Slack connected.

A strategist and an engineer open the same site. They see different accounts, because each signs in with their own permissions.

One cockpit. Everyone’s own permissions.
One cockpit. Everyone’s own permissions.

@ChatGPT in Slack

“Uploads over 20 MB fail on staging. @ChatGPT can you look?”

It checks Sentry, finds the pull request that lowered the limit and prepares a branch with a regression test. A client guest without a ChatGPT licence adds context in the same thread.

Nothing merges without review.

From a thread to a fix someone reviews.
From a thread to a fix someone reviews.

ChatGPT Space

Day two of a client workshop.

The Space already holds the goals, yesterday’s notes, the journey map and the prototype. Client guests are in it too.

“Turn today’s decisions into a page for the sponsor.” ChatGPT asks two questions first, then drafts while the room watches. The client comments on the page and tags @ChatGPT for the next round.

The readout is ready before the room empties. Client material stays where the client’s agreement says it belongs.

One room for the workshop. Before, during and after.
One room for the workshop. Before, during and after.

OpenAI Marketplace

NimbleFest 2027 needs a landing page.

We build it in Lovable, design it in Figma, run it from Notion and host it on Vercel. All four are Marketplace partners.

With an eligible agreement, part of an existing OpenAI commitment can pay for them. We still contract and pay each partner directly; OpenAI offsets the use.

A procurement detail that shortens the path from idea to a live page.

Built on our OpenAI commitment. Eligibility still needs confirming with OpenAI.
Built on our OpenAI commitment. Eligibility still needs confirming with OpenAI.

Also worth knowing

  • Dynamic pages, collaborative presentations, the Meetings plugin and shareable profiles move more of the work around the product into ChatGPT. For us, the question is where project context should live, within each client’s agreements.
  • A $500 Pro plan with the highest limits and Astra Ultrafast in ChatGPT Work and Codex.

Monday

Pick one useful thing your organisation already does.

Make it easier to get done.

Start with the person and the moment. Connect only what it needs. Decide what runs on its own, what needs approval and what always reaches an expert. Put the hard cases in the prototype early.

Think big. Start small. Ship to learn.

Which of these moments would you build first? Tell me, or reach out and we’ll map yours.

Join Bothrs

Bundle forces and let’s create impactful experiences together!
See positions

Agentic organizations don’t wait, they build.

Start your GenAI Discovery Track and unlock easier, smarter experiences.

Stef Nimmegeers, Co-Founder Nimble