Contents

Use dots for ongoing work across connected services, and start Codex with a clear goal and measurable completion criteria. Presentations at OpenAI DevDay Exchange Tokyo on October 20, 2026, showed how both approaches can move work forward while keeping people responsible for review.

Key takeaways#

  • dots can run cloud tasks on a schedule or in response to updates from supported services, and accepts new instructions while working.
  • Codex’s /goal lets an agent keep working toward a defined outcome within a budget and other stopping conditions.
  • Multiple agents help when tasks can proceed independently, but need clear responsibilities and shared constraints.
  • ChatGPT plugins connect conversations to external tools and interactive interfaces. Publishing requires checks and review.

Give dots a workflow with a defined scope#

At the Tokyo event, OpenAI introduced dots as an agent that researches, writes documents, and develops software based on users’ goals and past interactions. Tasks can start from direct requests, schedules, or updates from supported connected services.

In one development example, dots responded to a Slack bug report by investigating, changing code, testing, and creating a pull request. In another, dots was asked to migrate away from an API scheduled for retirement. Both are useful patterns for adoption because they produce work that people can review before accepting changes.

Before delegating, define the trigger and expected result. For bug fixes, specify the report source, target repository, required tests, and pull request requirements. Cloud work can continue with your laptop closed, but work using your local computer requires the ChatGPT app to stay open and online.

Keep shared deliverables in Spaces#

ChatGPT Spaces provides a shared place where people and agents can review pages, files, and deliverables. The presentation showed meeting notes being saved to shared pages, presentations being edited, and research being used to keep pages up to date.

The practical value is continuity. Recurring tasks need a consistent place for results so colleagues can find and revise the latest version. Without one, useful deliverables can end up scattered across conversations.

For your first workflow, choose one shared page as the destination and assign a human reviewer. Specify what the agent should update and what requires review before changes are made. Set information access permissions deliberately. Proactive research by dots depends on the information it is allowed to read.

Make Codex completion criteria measurable#

Bonnie Chatterjee’s session explained the Codex harness, the system that manages instructions, conversation history, tool execution, and results. For long-running tasks, context compaction reduces the information passed to the model while retaining what it needs to continue.

Retaining context alone does not guarantee completion. An agent may stop after reporting its progress. /goal addresses this by checking whether the goal has been achieved and allowing work to continue if it has not.

Pair an outcome with a way to verify it. If search is slow, specify the operation, baseline measurement, target, and permitted scope of changes. The official guidance cited in the report also describes limits: work may stop when the budget is exhausted, the user pauses it, or user input is needed to proceed.

Divide work only when it helps#

The session recommended starting with one agent. Adding agents increases coordination overhead and can duplicate work. It helps when a project includes tasks that can proceed independently, such as parts of a large refactoring effort or game implementation.

Before dividing the work, define four things for each agent.

  • The deliverable to produce.
  • The files or components it owns.
  • Constraints shared across the project.
  • Conditions that require a decision from someone else.

Maintain a shared work log recording the goal, current instructions, owners, unfinished tasks, and decisions. Agents should check it before resuming work after a handoff. The presentation also introduced an approach in which lightweight models handle routine work and more capable models help with difficult problems. Evaluate whether this setup meets your quality and cost requirements.

Connect services with review built in#

Corey Ching’s plugin session showed ChatGPT suggesting relevant integrations and displaying their interfaces alongside the conversation. Examples included a molecular viewer and a design tool for manually adjusting generated pages.

Each integration method serves a different purpose. MCP connects external data and tools, while interface extensions add controls and viewers. MCP Events can start work in response to updates from supported services. Choose the minimum integration needed to complete the workflow.

For distribution, the submission process presented requires uploading a ZIP package, resolving issues flagged by automated checks, and submitting it for review. After approval, developers choose when to publish. Sign in with ChatGPT was also introduced as a way to use a user’s ChatGPT plan allowance in supported external apps, but check eligibility and limits before relying on it.

Next step#

Run one Codex task with a defined scope. Choose a reproducible bug, define the tests that must pass and the files that may change, and request a pull request for human review.

Source: https://gihyo.jp/article/2026/10/openai-devday-exchange-tokyo-report

Contents