From Individual Agents to Team Workflows: A DevDay Demonstration

Jason of OpenAI’s developer experience team introduces a DevDay 2026 talk about moving from individual agent use to team collaboration. His examples describe an event demonstration and product direction; availability and behavior may change.

He says individual productivity can turn a person into the glue and bottleneck between tools. His definition of multiplayer is not merely a shared chat: it is a team system in which context, delegated work and decisions can move between people.

He introduces a hypothetical incident involving reports of rate-limit problems. The task is to gather context, determine whether issues are related and escalate appropriately without losing relevant information across browser and team conversations.

Earlier workflows relied on pinned threads and automations but still required him to coordinate everything. He says teammates could not easily pick up where he left off, so he remained the central point of contact.

He organizes the talk around bringing context into an agent, coordinating work and keeping shared information current. Context can come from a screen, browser, voice or connected services, subject to the access those tools have.

Jason demonstrates Appshots as a way to begin from the desktop context already visible, instead of manually describing every detail. The captured material becomes input to the agent, so users should understand what they are sharing.

He says capturing context alone did not remove the need to supervise every task. The next step in his example is a persistent agent that can continue an investigation and return with results.

He describes a cloud agent with its own browser and files, able to work while the user’s computer is off. This is a description of the demonstrated workflow, not a promise that all tasks will run unattended or succeed.

Voice lets him provide context quickly and informally. In his example, a conversation supplies possible causes, a remembered Slack message and requests to compare configurations and pull requests.

He contrasts this richer spoken context with a short request to investigate. The useful contribution is the information from the team; the agent still needs to check leads rather than treat every recollection as fact.

After meetings, he describes giving both teammates and an agent action items, including checking with colleagues and gathering feedback through connected tools. Such actions depend on the user’s instructions and the permissions granted.

The talk turns from collecting context to organizing it. Jason describes using shared pages and instructions so information remains readable, cited and consistently formatted as different tasks contribute updates.

He compares earlier Markdown and agent instruction files with page instructions and automations. The aim is a durable shared record that another person or agent can use, not a guarantee that generated citations are always correct.

Examples include tracking research experiments and coordinating DevDay work. He says delegated updates reduce his need to monitor every Slack message, while humans retain responsibility for decisions.

He describes plugins and extensions as shareable, versioned ways to package capabilities and instructions for a team. Updating a shared tool can spread improvements without each person rebuilding the same workflow.

Although the incident example is technical, Jason says he uses similar shared workflows for production, content planning and other team work. The underlying idea is to make the system useful across roles.

He closes by saying that automation gives him more attention for colleagues, research and creative work. His account is a personal experience from an OpenAI employee, rather than a measured productivity result for every team.