Skip to content

Agents: execution, threads & tools

How an agent run actually works — from waking up to running tools and posting results.

Capabilities

Agent threads & work types

Each agent run is a thread with a typed work type (conversation, role learning, task memory) and a state machine, with persisted run items (LLM calls, tool calls, context). An agent thread 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 threads stay aligned with the visible timeline via cursor-based delta sync; long task histories are compacted to stay within context.

Layered system prompts & scope policy

Behavior is shaped by layered developer prompts and scope policies (task assignee vs assistant, recurrence, role learning) injected with organization and conversation context.

Tool execution model

Agents act only through tools. Runs emit visible activity, some tools end the turn, and long-running tools (VM commands, Codex turns) execute durably so they survive across the run.

In-run how-to guides

Agents pull scenario playbooks (coding, Codex, recurring tasks, managing Rukies) into context before acting.

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 it with an optional handoff) — subject to review gates.

Active work surfacing

A "working now" view shows agents with in-flight threads, with cancel available to admins and collaborators.

Agent activity groups

Consecutive agent tool and reasoning activity collapses into compact, expandable timeline groups with live state (working, waiting for approval, worked, failed), so agents do not flood the feed. Codex sessions are grouped distinctly.

Last updated 2026-07-04