Agent work needs an operating system around it.
An internal iOS companion for running, reviewing, and resuming agent work across a software portfolio.
A task enters the runtime with its conversation and project context. The task should remain identifiable beyond a single connection.
Agent work continues independently of the phone connection. Closing the phone app or losing its connection does not stop an active task.
The mobile app shows conversations and fleet status, retains useful content through connection trouble, and reattaches when it returns.
Review outputs, request revisions, approve an action, or stop work. Recovery and undo make supervision practical.
Operating workflow, review experience, and agent-assisted implementation of the internal companion.
An internal TestFlight companion used with an actively running multi-product agent environment.
A disconnected phone should not lose the work.
The early companion tied a conversation to a screen and its network connection. Navigating away or backgrounding the phone made completed work difficult to recover. An agent can keep running while its interface loses the thread.
Separate the running work from the view.
Agent work continues independently of the phone connection. The phone lets me send requests, check progress, and review results. Closing the app or losing the connection does not stop the work. When I return, the app reconnects to the conversation and its results.
The interface groups the work into Chat, Fleet, Inbox, and Settings. Connection handling preserves useful content during a failure instead of replacing the whole screen with an error.
It runs beyond the demonstration.
Across the supporting runtime, 17 distinct specialist workflows had recorded activity in the 30 days through October 24, 2026. There were 1,464 recorded executions, excluding the frequent monitoring poller.
The work supports six live, public-facing products. Operating experience has been distilled into 25 reusable playbooks and 50 documented engineering lessons.
Use the provider that fits the work.
Bridge orchestration today is primarily Claude-based. Across individual applications I have implemented OpenAI Responses API integrations and constrained structured outputs, plus MCP interfaces, image generation, and speech-related workflows.
The goal is dependable work with inspectable inputs, outcomes, and human checkpoints.
Make review a normal part of the workflow.
Approval queues, revision requests, notifications, and undo make the internal system more usable. A real stop action and durable state matter when work takes longer than a mobile connection lasts.
The broader lesson is that an effective agent workflow includes task ownership, continuity, inspection, and recovery. The model call is one component in that operating system.