The short version
- Agent Optimizer, announced September 11, 2026 alongside the job-ready agents, is Salesforce’s tool for the least glamorous part of running agents: making them better after they ship.
- What Salesforce has said it does: work alongside teams across the agent lifecycle to build and refine agents, subagents, and actions, test performance, and analyze session traces.
- GA was slated for October 2026. As of this writing the switch has not publicly flipped; the Winter ’27 final production wave on October 10 is the window to watch. We will update this post when it lands.
- It fills a real gap: Testing Center tests what you ask it to, Session Tracing shows what happened, but the loop between them, reading traces and turning them into fixes, has been a human job.
- Pricing, credit consumption, and edition availability are all unannounced. The preparation work, baselines, traces, a regression set, pays off whether or not you ever turn Optimizer on.
The Loop Nobody Closed
Here is the unglamorous truth about running AI agents in production: shipping one is the easy half. After go-live, somebody has to read the transcripts, figure out why the agent escalated conversations it should have resolved, adjust instructions or actions, and verify the fix did not break something else. On most teams that somebody is a human with a spreadsheet, and the work happens in the gaps between other work.
Salesforce’s existing tooling brackets the problem without closing it. Testing Center runs the evaluations you define. Session Tracing records what agents actually did. But the loop between them, from "here is what went wrong in production" to "here is the refined agent that handles it," has been manual. Agent Optimizer is Salesforce’s stated answer to exactly that loop.
What Salesforce Has Actually Said
Agent Optimizer was announced on September 11, 2026, in the same release as the seven job-ready agents, and the official description is compact: it works alongside teams across the agent lifecycle, helping to build and refine agents, subagents, and actions, test performance, and analyze session traces. General availability was slated for October 2026.
That is the verified surface, and we are keeping this post honest by not decorating it. Pre-GA announcements in this space have a pattern: the capability list is directional, and the details that decide whether it works for you, what it can touch, what it consumes, where it runs, arrive with the GA documentation. The announcement names the lifecycle stages; it does not yet publish the mechanics.
Where It Sits in the Stack You Already Have
If you run Agentforce today, you already have pieces of this. Mapping them clarifies what Optimizer adds:
- Testing Center evaluates agents against test cases you author: pre-deployment quality gates, batch evaluations, regression runs you trigger.
- Session Tracing and analytics give you observability: what the agent did, step by step, in real conversations.
- Agent Script gives deterministic control for flows that must never improvise; our governance stack explainer covers where it fits.
- Agent Optimizer, per the announcement, is the connective layer: read the traces, refine the agents and actions, test the result. The loop, productized.
A caution worth stating plainly: until the GA documentation lands, any claim more specific than the announcement language, automatic configuration changes, self-running regression suites, autonomous tuning, is speculation, including when it appears in confident third-party roundups. We will revise this section against the real docs when they publish.
The GA Watch
October 2026 was the stated window, and the mechanics favor it: Winter ’27 reached most production orgs on October 3, with the final wave on October 10. As of this post’s publication, Salesforce had not publicly confirmed the GA flag. That is not a delay signal; it is just the normal gap between a stated month and a visible switch.
We are watching it, and this post gets updated the day GA is confirmed, with what the documentation actually says against each claim above.
How to Prepare While You Wait
The useful part: everything an optimizer needs is something your agent operation should have anyway. Three moves to make this month:
- Baseline your metrics. Resolution rate, deflection rate, escalation rate, and cost per resolved conversation, written down, per agent. Improvement claims mean nothing without a before. Our ROI playbook includes the kill-threshold discipline that makes these numbers honest.
- Keep session tracing on and retain the transcripts. Trace history is the raw material. If tracing is off today, the optimizer you enable in November has nothing to learn from.
- Curate a regression set. Twenty to fifty real conversations your agent must keep handling correctly, including the weird ones. Whether Optimizer runs them or Testing Center does, you need the set.
If you want help standing up that discipline, baselines, tracing, test sets, before the tooling arrives, that is a normal fixed-price scope for our Agentforce practice, and a free 30-minute call is enough to size it.
The Open Questions
Three things the announcement does not answer, which we will fill in at GA: price and metering (does Optimizer consume Flex Credits, and at what rate?), edition availability (unannounced, and note that no official per-agent edition mapping exists for the job-ready agents either), and surface (inside Agentforce Studio or a separate console?). If any of these decides your rollout, hold the decision until the documentation lands; it is days away, not quarters.