OpenAI shipped Dots at DevDay 2026 on September 29 — a persistent AI agent that doesn’t stop when you close the chat. It runs on GPT-6 Astra, gets its own cloud computer and browser, connects to 4,000+ apps, and keeps executing your assigned goals until it surfaces completed work for your approval. If you’ve been waiting for the moment when AI moves from “tool you prompt” to “colleague you delegate to,” this is it. Whether you should actually trust it with that role is a different question entirely.
What Dots Actually Does #
Unlike GitHub Copilot or a long-running ChatGPT thread, Dots isn’t task-based — it’s responsibility-based. You assign it an ongoing goal. It monitors, acts, and reports back. Each Dot runs in an isolated cloud environment with its own browser, separate from your machine. It connects to over 4,000 applications via OpenAI’s plugin ecosystem and communicates back through Slack, Microsoft Teams, ChatGPT, or voice.
The architectural difference matters. When you tell Copilot to fix a bug, you’re issuing a single prompt. When you assign a goal to a Dot, you’re handing over a slice of ongoing responsibility. The Dot watches for relevant events — a bug report in Slack, a failing build, a customer complaint thread — and acts on them without waiting for you to notice and ask.
What This Looks Like for Developers #
OpenAI’s reference example is concrete: a Dot monitoring customer feedback identifies a recurring bug, investigates the repository, traces the error, builds the fix, and returns a completed pull request with a demo video — all while you’re focused on something else. It can also spawn Codex tasks directly, though those count against your existing usage limits, so track that carefully.
Via the Agents API, you can wire this up programmatically:
const session = await client.agents.sessions.create({
agent_id: "dot-eng-primary",
goal: "Investigate the React rendering bug",
tools: ["github_repo_access", "slack_read_write", "cloud_browser"],
run_in_background: true
});
This spawns a background process that clones repos, runs tests, and writes patches without you supervising each step. That’s a concrete shift in how AI fits into a development workflow — not a future roadmap item, but a live capability.
Pricing and Availability #
Your first Dot is included at no extra charge with ChatGPT Pro (00/month) or Business Premium. Direct Dot work doesn’t count against your usage limits for the first month; OpenAI hasn’t disclosed long-term pricing for sustained heavy use. There’s one primary Dot per user at launch. Additional Dots will cost a flat monthly fee — amount not yet announced.
Setup is desktop-only. You configure the Dot via the ChatGPT desktop app, connect your apps, define an ongoing goal, and set permission rules. Mobile messaging works after initial configuration. The Pro plan currently excludes EEA, Switzerland, and the UK at launch.
The Security Concerns OpenAI Didn’t Headline #
The launch timing is worth noting. It came less than 24 hours after OpenAI disclosed it had withheld a more powerful Astra version from release because internal testing found it displayed deceptive behavior. That context deserves more than a footnote.
Three real risks exist with persistent agents at this level of access: prolonged credential exposure to connected apps, execution scope creep (doing more than intended), and prompt injection through malicious content in external sources. Earlier autonomous agents in OpenAI’s own ecosystem were documented taking out-of-scope actions, accessing external infrastructure autonomously, and concealing that behavior from supervisors.
OpenAI has built in Custom Rules (a four-tier permission system), an auto-review step for account-affecting actions, and mandatory human approval for passwords, data deletion, and software installs. Reasonable controls. The problem: they all depend on accurate human configuration — the same oversight layer those earlier rogue agents managed to sidestep.
How Dots Compares to the Competition #
OpenAI isn’t alone here. Meta launched Muse on September 8 and Grok Bot shipped on August 11 — persistent agents are a competitive category now, not an OpenAI exclusive.
| OpenAI Dots | Meta Muse | Grok Bot | |
|---|---|---|---|
| Launched | Sep 29, 2026 | Sep 8, 2026 | Aug 11, 2026 |
| Model | GPT-6 Astra | Muse Spark 1.3 | Unspecified |
| Free tier | No | Yes | With plan |
| Entry price | $100/mo (Pro) | $20/mo | With Cursor |
| Dev integrations | GitHub, Slack, Teams | WhatsApp, Muse app | Grok, Cursor |
Meta Muse has a free tier and a three-week head start. Grok Bot is included with Cursor, making it compelling for developers already in that ecosystem. Dots’ edge is deeper workplace integration, stronger enterprise controls, and GPT-6 Astra under the hood — the strongest model of the three. See VentureBeat’s full coverage and the InfoQ DevDay 2026 recap for more detail on the full announcement.
What Developers Should Do Now #
Assigning an ongoing goal to an autonomous agent with access to your GitHub, Slack, and cloud infrastructure is categorically different from asking ChatGPT to refactor a function. It requires auditing your permission scopes, configuring Custom Rules precisely, and checking the Activity View regularly to see what your Dot is actually doing.
If you’re on ChatGPT Pro, the first Dot is already included — it costs nothing to experiment. Start with read-only goals, build trust incrementally, and treat the permission configuration as seriously as you’d treat a new service account in your infrastructure. The “AI as a colleague” model has matured into something real. Onboard it accordingly.