Utah

Utah — short for Universally Triggered Agent Harness — is a conversational agent for Telegram or Slack with tools, memory, sub-agent delegation and full durability, built by Inngest and introduced in Your Agent Needs a Harness, Not a Framework. It is written in minimal TypeScript with no agent framework: Inngest functions, steps and events provide the harness around a standard think → act → observe loop. The company built it to show that durable, event-driven infrastructure can serve as an agent’s Agent Harness, and the post likens it to a durable, cloud-ready OpenClaw. Its source code is published as a reference implementation at https://github.com/inngest/utah.

“Universally triggered” means the agent does not know or care how it was activated — a chat webhook, a cron schedule, a sub-agent invocation or an inter-function event — so a new channel can be added without changing the agent loop.

Capabilities

Chat webhooks reach Inngest Cloud, where a webhook transform turns the raw HTTP payload into a typed event. A worker running locally, connected to Inngest Cloud over a persistent WebSocket without needing a public endpoint, picks up the event, runs the agent, and emits a reply event that a separate function sends back through the channel’s own API.

• The agent loop as steps. Each iteration calls the LLM, executes any tools it asks for, and feeds the results back; every LLM call and every tool execution is an Inngest step, so a failed call is retried on its own while earlier iterations stay persisted. A text response with no tool calls ends the turn.
• Six functions. The agent is split into a main loop, a reply sender, a typing indicator that fires immediately, a global failure handler, a scheduled heartbeat and a sub-agent function, each with its own retry policy, concurrency controls and triggers.
• Tools. File reading, writing and editing, shell execution and search come from the pi-coding-agent library; Utah adds remember (notes to a daily log), web_fetch and delegate_task.
• Sub-agents. delegate_task starts a separate sub-agent run with its own session key, context window and durability, using the same tools minus delegate_task so it cannot spawn further sub-agents, and returns a summary to the parent.
• One run per conversation. Singleton concurrency keyed on the session cancels the running loop when a new message arrives and restarts it with the latest context.
• Context management. Old tool results are soft-trimmed or cleared when context grows large while the last three iterations stay intact; the session history is summarized between runs once it passes a token threshold; the agent is warned as it nears its iteration limit; and a context-too-large error triggers forced Compaction and a retry.
• Model providers. LLM calls go through pi-ai, a provider-agnostic abstraction, so switching between Anthropic, OpenAI and Google is a configuration change.

Adoption & Ecosystem

Utah draws on OpenClaw and the pi coding-agent libraries, which the post names as its inspiration, but where those handle events and orchestration in memory, Utah hands orchestration to Inngest; the post credits that separation with traces and step-level inspection, retries, an audit trail of events, built-in scheduling, and groundwork for distributed, multi-player orchestration. As published it is a personal, single-player harness that runs on a local machine or a server. Its authors describe steering mid-run messages as unsolved and list making it multi-player, streaming progress updates, human-in-the-loop approval flows and letting it build new agents and workflows itself as things they are exploring.