September 2026 Summaries

Errand is an open-source macOS desktop application positioned as an alternative to Grok Bot and Muse, designed to let users delegate tasks to AI teammates that continue working in isolated cloud computers after the user closes the app. Its developer says the initial 4,000-line TypeScript client was built in about a week because it relies on Runta’s prebuilt infrastructure for persistent runtimes, browser and desktop control, agent execution, live VNC access, task persistence, event replay, artifacts, and API-managed state. Runta provides isolated machines with shells, browsers, and durable filesystems, while its Cloud Agents API maps directly to Errand’s user interface concepts such as teammates, tasks, activity, and files, leaving the app to focus on product decisions, desktop architecture, credential handling, attachments, packaging, and user-facing states. The author contrasts the rapid development of Errand with the roughly six months required to build Runta’s underlying platform, arguing that ready-made execution infrastructure allows developers to concentrate on the application experience rather than provisioning, job queues, event logs, and runtime lifecycle management. Errand remains early-stage, requires a Runta account and a user-provided model provider, supports swappable models and agent harnesses, and plans to offer a bring-your-own-cloud deployment option.

Errand is an Apache 2.0 open-source AI agent application designed to act as a persistent teammate that can complete tasks asynchronously on its own Runta-hosted computer, retaining files, state, and conversation context after a user closes the app. Users can assign tasks such as reviewing inboxes, repositories, Slack messages, and meeting notes, then later receive a summarized result that highlights relevant updates and decisions requiring attention. Runta supplies the underlying execution environment, including persistent storage, controlled service access, credential substitution that prevents models from directly seeing secrets, detailed tool-call tracing, token-usage monitoring, and checkpointing for recovery of files and processes. Unlike hosted agent products that rely on a provider’s infrastructure and model, Errand supports models and endpoints organizations already trust, including internal ones, targeting teams with stricter data-location requirements. Although Runta hosting is currently required, the developers are working toward a bring-your-own-cloud option, and invite users to access Errand’s source code on GitHub and build their own agent products with it.

Exo Harness is offered as a built-in Runta runtime that combines Exo’s agent framework with Runta-managed infrastructure, including model access, interactive terminals, background services, checkpoints, autoscaling, network controls, and secret management. Exo separates durable, protected infrastructure in the Harness from replaceable Executor logic for prompts, tools, memory, policies, and model calls, while an isolated Sandbox supports command execution and recoverable environment checkpoints; this design allows agent behavior to be modified without losing state, conversation history, artifacts, or secrets. Model credentials remain protected because Exo stores only secret stubs and Runta’s egress gateway injects API keys when forwarding requests to approved providers. Users can work in a resumable local-process environment through Runta’s web terminal, create an Exo runtime after adding a model-provider key to Runta Secrets, or use the Runta CLI skill to ask a coding agent to provision and verify a configured runtime.