Outcome graphs make agent work measurable

Production agents need process topology, evidence gates, durable execution, and fleet metrics, not just prompts.

Logged at IST: 2026-09-05 09:13 IST

What it is: Jeremiah Lowin expanding Mitchell Hashimoto's “outcome machine” framing into a short essay on outcome graphs for reliable agent workflows.

Gist: Lowin argues that making an agent succeed once is easy. Making repeatable, verifiable outcomes across a fleet needs a graph around the agent. Directed Agentic Graphs define measurable outcomes and permitted paths; agents stay autonomous inside nodes, while the system checks evidence at boundaries and controls what can run next.

The refund-agent example is the useful version: investigation can be flexible, but payment tools stay unavailable until eligibility and approval are verified. Cost, review burden, recovery, and ROI then attach to resolved outcomes, not just token spend or one-off demos.

Newsletter angle: This is the practical shape of agent productionization: autonomy inside nodes, hard checks at boundaries.