Appearance
Agents: execution, threads & tools
How an agent run actually works — from waking up to running tools and posting results.
Capabilities
Agent turns & work types
Each agent run is an agent turn with a typed work type (conversation, task memory, memory compilation, or assistant suggestion), a lifecycle state, and persisted activity. A turn can be processing, waiting for input or approval, blocked by a limit, or finished, cancelled, or failed. An agent turn is distinct from a conversation reply thread.
Agent worker runtime
Background workers wake on events, claim agents via leases, and run pending threads. Multiple worker instances share work safely, so a single instance failing does not stop the system.
Conversation-thread sync & compaction
Running turns stay aligned with the visible timeline via cursor-based synchronization. A fresh turn bootstraps from bounded useful conversation history, and long conversation or image-heavy contexts can be compacted into snapshots so the agent can keep working with manageable context. Provider image bytes and upload IDs are not persisted in the conversation.
Layered system prompts & scope policy
Behavior is shaped by layered developer prompts and scope policies (task assignee vs assistant and recurrence) injected with organization and conversation context.
Tool execution model
Agents act through registered tools. Runs emit visible activity, control-flow tools can end a turn, and long-running work such as VM commands and Codex sessions is tracked durably.
In-run how-to guides
Agents can fetch scenario playbooks for recurring tasks and managing Rukies when they need that guidance. Coding and Codex guidance is part of the fixed agent prompt.
Agent messaging & OpenUI
Agents reply in markdown with structured mentions, can intentionally abstain, and can render structured OpenUI widgets (for example, a home "mission control" briefing).
Task lifecycle tools
Agents propose a task plan, start work (setting the task in progress), and finalize (closing the task and sending a concise handoff to its receiving conversation) — subject to review gates. A handoff can link the detailed result message already posted in the task, preserving its generated artifacts with the result.
Active work surfacing
A "working now" view shows agents with in-flight threads, with cancel available to admins and collaborators.
ChatGPT subscription usage
A connected user can opt new agent turns and Home/Room memory jobs into their ChatGPT subscription. The setting is experimental and scoped to work started on that user’s behalf; other background processing continues to use workspace billing. A separate preference controls whether new VM Codex subscription usage contributes estimated API-equivalent cost to agent budgets.
Agent activity groups
Consecutive agent tool and reasoning activity collapses into compact, expandable timeline groups with live state (working, waiting for approval, worked, cancelled, or failed), so agents do not flood the feed. Codex sessions are grouped distinctly and show the recorded model and effort when available.