A long-horizon agent needs a definition of finished
Memory and durable execution can keep an agent working for weeks. They do not define when the business outcome is complete, accepted or uncertain.
Salesforce has given its new agents names, jobs and a runtime that can pursue goals for days or weeks. Hunter works a sales pipeline; Paige handles employee requests; Marshall runs back-office processes. That is a stronger product boundary than “general assistant”. It still leaves the most important field blank: what evidence proves the job is finished?
My rule is simple: a long-horizon agent needs a completion contract before it needs a manager persona. Memory, durable execution and steering can keep activity moving. None of them decides whether the business outcome was accepted, failed or left uncertain. This is my editorial judgment, not a control I have deployed in a Trigent or client Agentforce implementation.
Salesforce's six results do not share one denominator
Salesforce's 11 September 2026 release reports six customer outcomes. Engine has 50% of chat inquiries fully resolved. Perk says Hunter builds 60% of its sales pipeline. Autism Queensland has 70% of administrative requests resolved by Paige. Hibbett AI handles 90% of core shopper journeys, Asana reports four times the conversation volume, and 79% of Anthropic conversations seen by Fin are resolved autonomously.
These are Salesforce-reported marketing results; I could not verify their denominators, baselines, measurement windows or causal attribution. More importantly, they measure different things. A resolved request, pipeline contribution, handled journey and increased conversation volume cannot be rolled into one “digital labour productivity” score.
That mismatch is useful. It shows that the job definition must choose the outcome. A service agent may finish when the user's issue is confirmed resolved. A sales agent may finish one step while the opportunity remains open for weeks. More activity can be progress in one job and unfinished inventory in another.
Durable execution extends work; it does not accept it
Salesforce says Hunter is the first agent on its long-horizon runtime. The product combines memory across sessions, durable execution that can resume or change course, and dynamic steering from a user. The release also says Hunter is in pilot, with general availability planned for November 2026; the wider promise should not be read as current availability for every named agent.
Those three capabilities solve continuity. They do not supply acceptance criteria. An agent can remember the plan, resume after an interruption and respond to feedback while still optimising the wrong outcome.
Salesforce's Agent Script page adds deterministic workflow rules, persistent variables, simulation and trace data. These are valuable controls. A trace can show exactly which actions happened, but an accurate activity record is not the same as evidence that the requested business result happened.
Write the completion contract before the first run
For a hypothetical outbound-sales agent, I would define six fields before authorising a multi-week run: the business outcome, deadline, acceptable evidence, actions allowed without approval, mandatory review points and a terminal uncertain state. “No qualified reply by Friday” must be a legitimate result. Otherwise the agent is rewarded for producing more outreach because silence has no representation.
The same contract should separate the agent's part from the whole job. Drafting a renewal proposal may be complete when the approved document exists. Retaining the customer is not complete until the customer decides. Calling both outcomes “done” turns a useful worker into a generous attribution machine.
This advances an older AndyMental rule: agent count is not a production metric. Once the jobs are named, the next question is not how many agents own them. It is how each job closes without confusing activity, contribution and accepted outcome.
What's in it for you
- Compare agents on accepted outcomes inside the same job, not across incompatible vendor percentages.
- Give every long-running goal explicit failed and uncertain endings, not only a success path.
- Review the completion evidence before the trace; use the trace to explain the result, not substitute for it.
A long-horizon agent is ready for work only when the system knows what finished, failed and uncertain mean.
