Which roles keep appearing in human-agent teams?
Three show up repeatedly once agents are real: someone who improves the agent by feeding it organizational context, someone who handles the exceptions it escalates, and someone who coordinates several agents and resolves conflicts between them. The pattern worth noticing is that all three are human jobs created by automation, not jobs removed by it, which is usually the more accurate framing for a mid-market plan.
How should work be divided between people and agents?
Split by consequence, not by difficulty. Give the agent the volume and keep the judgment, then make the boundary explicit: what triggers escalation, what context travels with it, and what the person is expected to decide. That boundary is a real design artifact, the same discipline as an agent handoff between agents, and it fails the same way when context is dropped. Keep a per-decision control through human in the loop, name an owner as part of the agent lifecycle, and expect the teaming model to need AI literacy in the receiving team before it works.
How do we divide work between people and agents at Soba Labs?
We run the company with agents in the middle and people at the edge, and every action an agent can take sits in one of three tiers. It acts alone and records the action: reading, writing to our own records, opening a pull request, drafting a message, running tests. It does the whole job and stops at a reviewable artifact, for a person or an automated gate to decide, when the next step is a merge, most outbound messages, a publication, or a commitment to a price, date or scope. Or it asks first and waits for a yes on that item: anything on a person’s machine or infrastructure, spending money, a production data write. The tier belongs to the action, never to how confident the agent is.
Agents that run unattended on a schedule get the same tiers. Three rules keep the line: a question is not a yes, a yes covers one item in one session, and an unclear tier means ask. Autonomy grows only when a person deliberately moves a class of work into the act-alone tier and records why. Writing these decision rights down in one place is what lets an unattended agent see the line at all.
How do you know when an agent has earned more autonomy?
From agreement data, not from a falling override rate. When people shift from correcting to skimming, overrides drop through inattention, so pair any override metric with blind re-review of a random sample of the work the agent approved alone. Run the agent in shadow mode beside people first, measure agreement per segment, then stage the rollout: suggestions only, then low-risk actions with approval, then automatic action with an audit trail.