The most quietly radical thing Salesforce announced this cycle is not a model. It is a runtime. Hunter, currently in pilot with general availability planned for November, is the first agent built on a new long-horizon runtime — meaning it can hold a goal across days or weeks instead of resetting after every interaction.
Alongside it sits Koa, Salesforce's own reasoning model, post-trained from NVIDIA's Nemotron 3 Super on synthetic data drawn from nearly 30 years of CRM deployments across 14+ industries. Salesforce claims Koa makes three times fewer errors than leading models on its own CRM benchmark — a vendor benchmark, worth noting, not an independent one. Koa is in pilot now with general availability expected this winter in US regions.
Put the two together and the shape of 2027 agent work is visible: agents that pursue a goal for weeks, reasoning with a model tuned on CRM history. The question nobody's keynote answered is the one that decides whether this works in your org: what does supervision look like when the work outlives the session?
Why session-based review stops working
Every agent oversight pattern in production today assumes the session is the unit of work. A human reads the draft, approves the action, closes the loop. That pattern has a hidden dependency: the work is short enough for one person to hold in their head at the moment of approval.
A long-horizon agent breaks that dependency three ways:
- The goal outlives the reviewer's attention. Nobody re-reads week-one context when approving a week-three action.
- Small drifts compound. A slightly wrong assumption on day 2 is an invisible premise by day 12.
- The approval moment disappears. When an agent acts continuously, there is no natural pause where review happens — unless you build one.
Supervision patterns that survive long horizons
- 01Checkpoints on the calendar, not on completion. A long-running agent gets reviewed on a schedule — not when it finishes, and not only when it asks. The review covers trajectory, not just output: is the goal still the right goal?
- 02Budgets as hard limits. Actions per day, records touched, spend, messages sent. A long-horizon agent without budgets is an unbounded liability with a login.
- 03Resettable premises. Record the assumptions the agent is operating on, and make them reviewable. When reality changes — a pricing update, a policy change, a client escalation — someone must be able to invalidate a premise and force a re-plan.
- 04A kill path that works mid-flight. Stopping an agent three weeks into a goal must be a clean operation with a known blast radius, not an incident.
- 05Human-tested before customer-facing. Vendor benchmarks measure a model. They do not measure your org's data quality, your edge cases, or your escalation paths. Every long-horizon workflow needs a supervised dry run on real work before it touches a customer.
The uncomfortable part
Long-horizon agents will be sold on headcount math — one agent pursuing pipeline for weeks, no salary. The math only holds if someone owns supervision, and supervision of long-horizon work is a real job with real hours. Budget for it up front. The alternative is discovering in week four that nobody was watching weeks one through three.
Where we stand
We have been explicit about this stance: AI does the work, humans test the work, and nothing ships to a customer surface untested. We run agents on our own operations under exactly these rules. Our 69 consultants carry 221 Salesforce certifications — 56 of them Agentforce — and across 45+ contracts since 2019 the lesson repeats: the constraint on agent value is never the model. It is whether the organization around the agent knows what the agent is doing.
Hunter goes GA in November. If you want your supervision model designed before your first long-horizon agent starts its first long horizon, that is a conversation worth having now.