Greg Isenberg watched OpenAI DevDay 2026 through a founder's lens. His shortlist is not every model announcement: it is the combination of an always-on personal agent, a narrow decision service, a managed execution layer, and a way for eligible users to bring their ChatGPT identity and plan to outside software. Those pieces could make small, specialized workflows easier to distribute and operate. They do not, by themselves, create a business.
Watch Greg's DevDay Breakdown
Credit: Greg Isenberg's Startup Ideas Podcast episode, published 29 September 2026. The two startup concepts, four-step framework, and market forecasts are Greg's analysis. Product status below is checked against OpenAI's public documentation as of 30 September 2026.
Four Products, Four Different Roles
| Development | Job in the stack | What a builder should verify |
|---|---|---|
| Dots and plugins | A personal agent may use connected capabilities to complete a user task | Eligibility, plugin review, selection, and completion rates |
| Decisions API | Fixed-choice classification or routing, as described in the episode | Preview access, output contract, latency, accuracy, and price |
| Agents API | Managed agent sessions, tools, subagents, and hosted computer use | Approval boundaries, environment cost, recovery, and audit logs |
| Sign in with ChatGPT | Identity sign-in and, separately, eligible plan-backed AI requests | Partner approval, user consent, model scope, limits, and fallback billing |
Do not collapse these into one launch. A dot is a user-facing agent, a plugin is an external capability, the Agents API runs an application-owned agent, and sign-in is an identity and authorization flow. A company could use one without using all four.
Dots Are a New Surface, Not Guaranteed Traffic
Greg's thesis begins with a customer asking an agent to finish a job: find a restaurant, resolve a permit, or coordinate a service. If a plugin can supply the missing action and a reliable result, it may become relevant in the flow before the user searches for the vendor by name. That is a compelling distribution hypothesis, but it is still a hypothesis for a new builder.
OpenAI's Dots guide describes an always-on agent with its own cloud computer and connected apps. It also says access is rolling out gradually and the first dot is included in eligible Pro or Business Premium plans. A dot's availability is not equivalent to every ChatGPT account seeing every plugin.
For plugins, OpenAI's review requirements distinguish publication from enhanced distribution. A published plugin can be found by exact name or directory link; prominent placement and proactive suggestions are selective, not something developers can demand. Metadata guidance recommends testing direct prompts, indirect requests, and negative cases to see whether a tool is selected appropriately. The practical moat is the task you can complete, not an assumed share of a very large audience.
Decisions API: Interesting Preview, Unproven Economics
At 05:02, Greg describes the Decisions API as a limited-preview service that focuses GPT-6 Luna on user-defined questions with finite answers. Inputs could include text or images; outputs could route a support request, classify content, or select an agent's next action. He compares the idea with TypeSafe's Jev, but does not claim the two have the same performance.
The useful product question is narrower than "Can it decide?" It is: for a specific decision, is its error rate, speed, and cost better than a rule, a classifier, or a general model? For Greg's proposed UI example, a decision layer might rank candidate app-store screenshots or landing-page variants, but actual conversion lift needs a real experiment. A model's score is not a conversion result.
Agents API: Execution With Reviewable Boundaries
OpenAI introduced the Agents API on 10 September 2026, before DevDay, in public beta. It offers a managed Codex-style harness with long-running sessions, context compaction, tool search, and subagents. The newer computer-use guide shows an agent operating an OpenAI-hosted browser while the application handles origin access, authentication, and review.
Greg sees the opening in vertical expertise: give the managed agent your domain-specific data, tool permissions, and workflow design. A permit service, for example, might assemble requirements, prepare a draft, and route it to a human expeditor. The agent infrastructure reduces how much harness code a small team writes; it does not remove the need for supplier agreements, security controls, exception handling, or human approval before consequential actions.
Sign in With ChatGPT Has Two Separate Permissions
Greg's boldest economic argument is that a customer could use an existing ChatGPT plan to try a niche product, while the builder charges for team features, proprietary data, or a completed workflow. OpenAI's quickstart supports a narrower version: identity sign-in is one capability; eligible ChatGPT plan usage for AI requests is a separate permission. Plus and Pro users may use their plan in participating apps, subject to available credits and authorization.
At launch, commercial sign-in partners are in a limited trial, and plan usage is available to open-source partners and selected private clients. The user must approve the relevant scopes. Identity access does not reveal their conversations, API key, or billing details. An ordinary commercial app should therefore not assume that a free AI core is suddenly costless. First check eligibility, what requests qualify, what happens when a plan limit is reached, and how the product pays for everything beyond inference.
The enduring business idea is still valuable: sell a specialized outcome rather than a generic token wrapper. Greg lists CAD-file cleanup, a focused scientific research monitor, Shopify catalog maintenance, a repeatable editing workflow, and a narrow contract-review product as examples. Each needs its own dataset, operational workflow, expert review, and willingness-to-pay test.
The Trigger, Decision, Action, Feedback Loop
Greg's framework is a useful way to avoid building a plugin with no job to do. The four steps describe a complete service, not four mandatory OpenAI products:
| Step | Invoice example | What to own or measure |
|---|---|---|
| Trigger | An invoice reaches its due date | Permissioned event source and the right timing |
| Decision | Determine whether to remind, ask a person, or wait | Rules, evidence, confidence, and an escalation path |
| Action | Prepare a reminder or update the account record | Narrow tool scope, approval, idempotency, and rollback |
| Feedback | Track approval, payment, rejection, and customer response | Outcome data that improves the next workflow |
His thesis is that the most defensible businesses own the trigger, the real-world action, and the feedback, while a general assistant may handle much of the decision. That is a strategy, not a guarantee: control over data and fulfillment still has to be earned from customers and partners.
Two Business Ideas From the Episode
1. An API for Real-World Work
Many requests end outside a browser. A permit expeditor, customs broker, or other vetted specialist must do the last mile. Greg proposes an API that lets an agent request a bounded service, receive a quote, secure user approval, assign the job, and report status. The possible revenue is a service or transaction fee, not a fee for having an endpoint.
Start small: one service category in one jurisdiction, ten providers, one structured request, manual matching, and a visible status trail. Measure quote accuracy, acceptance rate, completion time, disputes, and repeat use. The hard work is verifying suppliers and taking responsibility when the physical-world job fails.
2. Analytics for Agent Discovery
If plugins become a channel, builders will need to know whether their tools are selected for relevant user intent and whether invocation leads to a completed task. Greg imagines a testing layer that generates indirect prompts, tracks tool selection and activation, diagnoses metadata gaps, and replays tests after every plugin update.
Start small: collect a labeled set of direct, indirect, and negative prompts for one plugin, then compare expected versus actual tool calls in a development environment. OpenAI already documents this golden-prompt method. A commercial analytics product would need to add cross-version history, completion measurement, privacy-safe reporting, and evidence that its recommendations increase successful tasks. It cannot assume access to other platforms' private ranking or selection logs.
A Sensible First-Week Test
- Choose one recurring request where a customer can name a successful outcome and its current cost.
- Map the trigger, decision, action, and feedback, including the point that requires explicit human approval.
- Manually fulfill five examples before building an agent. Record time, exceptions, and what customers actually value.
- Prototype one narrow tool and test direct, indirect, and negative prompts. Track selection and task completion separately.
- Check current product access and permissions before promising a plugin listing, Dots reach, Decisions API speed, or ChatGPT-plan-backed usage.
The opportunity is not a prediction that every niche app will find free distribution. It is that a smaller company can now test a complete, specific workflow with less infrastructure work. A real customer result is the unit worth validating.
Video Chapters
| Time | Topic | Time | Topic |
|---|---|---|---|
| 00:00 | Why DevDay matters to founders | 01:48 | Dots |
| 03:29 | Plugin comeback | 05:02 | Decisions API |
| 06:43 | Agents API computer use | 08:01 | Sign in with ChatGPT |
| 12:36 | Where to build | 13:46 | Real-world work APIs |
| 15:04 | Agent-discovery analytics | 17:05 | Closing thoughts |
Sources and Link Map
- Creator analysis: Greg Isenberg's full episode and the official DevDay event page.
- Dots and plugins: Dots access and setup, plugin review and distribution, and metadata and discovery testing.
- Agent execution: Agents API launch and hosted computer-use guide.
- Account access: Sign in with ChatGPT quickstart and identity and data-sharing help.
For a comparable third-party-agent platform thesis, read our Muse connector breakdown. For the fast-decision layer that Greg compares with this launch, see nine Jev use cases.