Type: url
Original: https://www.inngest.com/blog/your-agent-needs-a-harness-not-a-framework
Status: partial
Summary
An Inngest blog post by Dan Farrelly (3 March 2026) arguing that an agent runtime needs a harness — the layer that connects, protects and orchestrates the LLM, its tools and its memory without doing the work itself — and that durable, event-driven infrastructure already provides one. It presents Utah (Universally Triggered Agent Harness), a reference implementation in which every LLM call and tool call is an independently retryable Inngest step, the agent is split into six event-connected functions, sub-agents run via step.invoke(), one run per conversation is enforced with singleton concurrency, and context is kept in check with two-tier pruning, compaction and overflow recovery; it names steering and streaming as open problems.