Salesforce expands Agentforce with Coworker and long-horizon agents
The Dreamforce roundup spans GA, beta, pilot and planned capabilities. RevOps should govern each release state and revalidate long-running work as customer context changes.
Salesforce separates several Agentforce release states
Salesforce's September 28 Dreamforce roundup describes Agentforce Coworker as generally available. It says Coworker can answer natural-language questions across CRM, Slack and connected sources, build a plan and execute it while orchestrating other agents inside the conversation.
The same roundup describes long-horizon agents as pilot with general availability planned for November 2026, Agent Optimizer as GA in October and AI Skills as pilot with October GA. RevOps should preserve those distinctions. A broad Agentforce label is not a reliable statement of what is production-ready in a particular org.
Cross-surface access does not transfer authority
Salesforce says AIforce brings Salesforce to Claude, Slack and Salesforce itself under defined permissions, with different availability for Claudeforce, Slackforce and Coworker. Carrying context and governance to another surface can reduce switching, but the external surface still needs the correct user, session, object and field permissions.
Test least-privilege roles rather than an administrator account. Confirm what context is retrieved, what actions can execute, where approval appears and how evidence returns to Salesforce. A conversational surface should not become a route around CRM sharing, field-level security or change control.
Long-horizon work needs expiring approval
Salesforce says long-horizon agents can pursue goals over days, weeks or months, with memory, durable execution, conversational steering and approval check-ins under defined guardrails. That duration creates a different control problem from a single-turn assistant.
Customer, owner, opportunity, preference and contract state can change while a plan remains active. Require a current-state check before every consequential step and expire approval when the source purpose, owner or commercial facts change. Store which plan version and evidence justified each action, not only the goal that started the run.
Durable execution must include durable cancellation
A runtime that can continue after failures needs explicit pause, cancel, retry and compensation behavior. Operators should know which step is active, which external actions are pending and which completed actions cannot be reversed. Steering a plan should create a new version when it changes material scope.
Test a connector outage, permission revocation, stale record, changed owner, closed opportunity, duplicate task and delayed response. After cancellation, enumerate work already emitted to email, CRM, support or other systems. Do not assume stopping the plan retracts terminal effects.
Optimizer changes are production changes
Salesforce says Agent Optimizer can build subagents and stub actions, inspect production sessions, find recurring failure patterns, rank them by impact and help build, test and stage changes. The described workflow can shorten investigation and iteration.
Keep diagnosis, proposed fix, test, approval and deployment separate. Preserve the sampled sessions and failure definition behind a recommendation. A ranked pattern does not prove root cause, and an automatically generated action should not reach production merely because the same system suggested the repair.
Reusable skills expand the blast radius
AI Skills are described as reusable, governed instructions that can be built once and used across agents. Reuse can improve consistency and reduce duplicated configuration. A skill change can also affect several workflows, teams and surfaces at once.
Maintain dependency visibility, semantic versioning, test fixtures, owners and a rollback or pinning strategy. High-impact skills should preserve input and output contracts and fail closed when required context is missing. Review every consumer before broadening data or action scope.
Testing and monitoring cover different questions
The roundup describes batch prompt testing, voice and conversation testing, custom scorers, A/B experimentation in beta, Voice Observability and Agent Health Monitoring. These surfaces can expose reliability, resolution, latency and session-quality signals.
No single score proves that an action was authorized or that the final CRM and customer state is correct. Define local tests for normal and changed state, preserve scorer versions and independently read material destinations. A health alert should open an investigation; it should not automatically rewrite the agent without a controlled release.
Deterministic output is not deterministic business state
Salesforce also describes deterministic rendering, where the same structured response renders consistently across clients. Consistent presentation is valuable for forms, cards and other interaction surfaces. It should not be confused with deterministic decision or data quality.
The structured response can faithfully display stale, incomplete or unauthorized content. Preserve source evidence, object IDs, evaluated-at time and permissions behind the response. Test the customer or employee action that follows the interface, not only the pixels.
What RevOps should do now
Create a capability register that lists Coworker, long-horizon agents, Optimizer, AI Skills, testing and monitoring separately with official status, org availability, permissions, owner and last verified date. Select one bounded Coworker workflow and keep pilot or future capabilities out of its required path.
Test normal, stale, changed-owner, revoked-permission, connector-failure, retry, steering and cancellation cases. Require independent final-state reads and a named recovery owner. Salesforce's announcements widen the agent operating surface; disciplined release-state and change control are what keep that surface understandable after launch.
Original source
This DailyRevOps article is written in our own words from the source signal and adds RevOps context, workflow analysis, and operator interpretation.
- Original source: Salesforce
- Original publication date:
- Source link: Read the original article
Salesforce published the official Dreamforce 2026 roundup on September 28, 2026. DailyRevOps first published this operator analysis on October 1, 2026.