OpenAI Dots: Always-On Agent Coworkers, Connected Apps Architecture, and Enterprise Security Analysis

Persistent Autonomy: The Shift from Chat to Always-On Coworkers

At DevDay 2026, OpenAI introduced Dots, a platform that transforms AI interactions from reactive prompt-response sessions into persistent, autonomous coworkers. Visually embodied by customizable, fluidly animated character avatars, Dots run continuous background event loops across enterprise and personal tool stacks.

While previous chat-based interfaces required active human input to initiate each task, a Dot maintains a persistent state vector across days and weeks. It monitors incoming webhooks, processes batched tasks during low-usage hours, and presents completed deliverables upon the user return.

Core Architectural Components

The Dots runtime relies on three primary structural subsystems:

Dots Agent Architecture
├── 1. Unified Event Bus & Scheduler
│   ├── Webhook listeners (GitHub, Linear, Slack, Gmail)
│   ├── Cron task schedulers (Morning briefings, nightly CI runs)
│   └── Event deduplication and priority queues
├── 2. Isolation & Execution Sandbox
│   ├── Firecracker microVM execution environments
│   ├── Ephemeral Linux containers with bash and python runtimes
│   └── Read-only filesystem mounts with encrypted credential vaults
└── 3. Policy Enforcement & Guardrail Engine
    ├── Scoped OAuth token broker (least-privilege permissions)
    ├── External communication rate limiters
    └── Approval gating (Push alerts for high-consequence actions)

1. Persistent Memory Vector: Dots maintain structured entity graphs containing team hierarchies, project timelines, preferred formatting rules, and ongoing ticket references.
2. Action Broker: An abstraction layer that maps model tool calls to verified third-party API mutations.
3. Approval Gating: Actions exceeding predefined risk thresholds (such as emailing external clients or deleting repository branches) halt execution until confirmed through the OpenAI mobile companion app.

Enterprise Connected Apps Integration

OpenAI built deep connectors into five foundational enterprise software suites:

Security, Isolation, and Human-in-the-Loop Safeguards

Security researchers and enterprise CISOs raised valid questions regarding always-on agents operating with access to company data. OpenAI addressed these concerns in its official security documentation:

1. MicroVM Sandboxing: Every background execution runs inside an isolated microVM with strict network egress policies. The agent cannot scan internal VPCs or access peer containers.
2. Cryptographic Action Gating: Read actions occur automatically according to user-granted OAuth scopes. Write operations are tiered into safe actions (commenting on a ticket, drafting a branch) and critical actions (production deployment, external emails), which require human biometric signature.
3. Data Segregation: No customer data, conversation transcripts, or repository code processed by Dots are used to train foundational OpenAI models.

Competitive Dynamics: OpenAI Dots vs. Meta Muse

The introduction of Dots positions OpenAI in direct competition with Meta Muse, introduced at Meta Connect 2026:

OpenAI focus remains centered on verifiable cognitive labor: writing code, balancing schedules, generating technical reports, and triaging communications.