Autonomous software agents can now run continuously across digital workflows

Continuous digital assistants that operate without waiting for instructions could reshape automated software engineering and enterprise workflows, according to new releases from OpenAI.

Software automation has long operated on an explicit trigger, requiring a person to provide an instruction, review the immediate output, and issue the next prompt. When an unexpected error occurs or a digital environment shifts, traditional automation halts until a human intervenes to restart the process. Autonomous software agents alter this arrangement by keeping a persistent operational loop alive. Instead of vanishing between prompts, a persistent agent maintains memory of its goals, checks its intermediate outputs against expected standards, and decides which steps to take next without human prompting.

Moving from intermittent prompting to continuous execution changes how digital work gets done across computing systems. For automated tasks to proceed smoothly across hours or days, an underlying machine learning model must handle extended context, make decisions across external software interfaces, and avoid compounding errors that derail complicated operations. During DevDay 2026, OpenAI announced more than 20 updates across its models, ChatGPT, Codex, application programming interfaces, and enterprise products to support autonomous operation.¹

How do persistent software agents operate independently?

Persistent agents operate by keeping a continuous cycle of observation, planning, action, and adjustment running across digital environments. In this process, the agent receives an overarching objective, inspects the current state of a computer desktop or browser interface, selects an action such as clicking a menu or modifying a line of code, and evaluates the resulting change. If the action produces an unexpected error, the agent reviews the feedback from the software environment and chooses an alternate path forward.

OpenAI introduced Dots, which are always-on artificial intelligence agents designed to keep working independently instead of waiting for another prompt.¹ The company noted that these systems closely parallel Meta's recently launched Muse.¹ According to OpenAI, Dots can run 24 hours a day and seven days a week, learn from user feedback, operate their own computer and browser, and connect to more than 4,000 external services.¹ The company reported that Dots are powered by its GPT-6 Astra model and will eventually support voice interaction, Slack, and other messaging applications.¹ Alongside these agents, OpenAI announced Spaces, a collaborative digital workspace designed to host Dots, documents, presentations, Living Pages, and shared content management.¹

What performance gains make longer automated tasks feasible?

Executing long-running tasks requires underlying models that maintain low error rates over extended context windows while remaining computationally affordable. OpenAI released GPT-6.1 Sol as an upgraded version of GPT-6 Sol designed specifically for agentic coding, computer use, and long-running workflows.¹ According to OpenAI, the model approaches GPT-6 Astra performance at about one-fifth of the token cost.¹ GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens, operating with an expanded context window of 1 million tokens.¹

OpenAI reported that GPT-6.1 Sol matched GPT-6 Astra on DeepSWE 1.1, a benchmark designed to evaluate software engineering capabilities.¹ On the OSWorld 2.0 benchmark, which measures how effectively an artificial intelligence model navigates operating system interfaces, GPT-6.1 Sol improved performance by 7% over the earlier baseline.¹ For multi-step tasks where small mistakes can derail an entire automated sequence, OpenAI stated that GPT-6.1 Sol reduced factual errors from 11.4% down to 7.7%.¹ The company also stated that an Ultrafast version tailored for Codex is planned.¹

What additional developer tools support autonomous systems?

Developer tooling has expanded to give autonomous programs dedicated runtime environments and specialized decision interfaces. OpenAI announced that its Codex programming environment gained dedicated cloud environments for long-running tasks, voice controls, an /agents view, and Codex Security Cloud, which scans GitHub repositories and generates security patches.¹ To coordinate these capabilities across external software, OpenAI expanded its Agents API to include computer use and multi-agent coordination.¹

OpenAI also introduced the Luna-powered Decisions API to handle focused routing and classification tasks.¹ For commercial communication platforms, the company added @ChatGPT integration in Slack and Microsoft Teams.¹ It introduced ChatGPT Pro 500, a $500 monthly subscription tier that includes Astra Ultrafast, and introduced Sign in with ChatGPT support to let subscribers apply paid plan capacity across 16 third-party tools.¹ For organizations subject to stringent data governance rules, OpenAI previewed Private Inference.¹

What are the limits of autonomous agent systems?

These benchmark results represent synthetic evaluations and commercial vendor announcements rather than independent third-party performance verifications. The 7% performance improvement on OSWorld 2.0 and the reduction of factual errors to 7.7% show progress on standardized digital tests, but benchmark scores cannot guarantee reliable performance across arbitrary, messy real-world corporate codebases.¹ Operating agents continuously across browsers and thousands of external services also introduces unpredictable risks if a system encounters ambiguous security prompts or misinterprets instructions. OpenAI has not published comprehensive third-party audit data detailing failure recovery when Dots operate entirely without human supervision.

For organizations adopting autonomous tools, persistent operation shifts the role of human workers toward oversight and goal definition. When routine administrative tasks, code patches, and software integrations occur continuously in the background, engineers spend less time writing individual instructions and more time establishing boundary rules and auditing agent decisions. The next phase of deployment will test whether always-on agents can sustain complex, multi-day assignments without drifting from their initial objectives.

What this rests on

20 sentences trace to 1 source.

All 1Documents 1Records 0Reporting 0People 0

1. 1 A alternativeto.net Document · alternativeto.net · 1 Oct 2026 OpenAI launches GPT-6.1 Sol and Dots, its new always-on AI agents that work independently See the source