Open vs. closed agents: two modes, two contracts

Every agent we deploy runs in one of two modes, and the distinction matters more than the model behind it.

Open agents: exploration with a human in the loop

An open agent works with someone. It proposes, the human steers, it executes the next step. Think of drafting a quarterly narrative, investigating why churn spiked in one segment, or restructuring a campaign. The task is fuzzy, the definition of done lives in someone's head, and the value comes from speed of iteration.

The contract for an open agent: it never acts beyond the current step without confirmation, it shows its sources, and it leaves a transcript. Trust comes from visibility, not from autonomy.

Closed agents: scheduled, autonomous, self-checking

A closed agent runs without anyone watching. Monday 06:00, the reporting agent pulls the numbers, builds the summary, runs its eval pass, and delivers — or escalates. The task is precisely specified, the definition of done is written down as checks, and the value comes from never having to think about it again.

The contract for a closed agent is stricter: explicit input sources, deterministic output format, a written eval suite it must pass before anything leaves the system, and an escalation path with a full trail when it fails twice. No eval suite, no autonomy. That rule is not negotiable in our deployments.

The failure mode: shipping an open agent as a closed one

Most "AI automation went wrong" stories reduce to one mistake: a workflow that needed a human in the loop was scheduled to run without one. The agent was fine; the contract was wrong. The inverse mistake is quieter but expensive — keeping a human in a loop that checks nothing, signing off drafts they stopped reading in week two.

The practical sequence

Start every new workflow as an open agent. Watch where the human actually intervenes for four weeks. The steps where nobody ever corrects anything become candidates for closing — with the human's silent judgment converted into written checks. Closing an agent is not flipping a switch; it is graduating a workflow, with the eval suite as the diploma.