Ema has outlined a system it says can help its workplace AI agents improve from past decisions, while keeping proposed changes away from live work until an administrator approves them. The notable idea is not just learning from past attempts, but rehearsing changes against recorded evidence before they are applied.
Ema Watch analysis
What happened
In a post published on 19 November, Ema describes Dream RSI, a process for turning the work and feedback of its AI Employees into proposals to improve workflows, model choices and policies. Ema says a separate background process reviews decision records, such as approvals, corrections and later outcomes, to find recurring problems and suggest changes.
The company describes three ways to test proposals: check a proposed rule against recorded decisions, replay past requests using stored inputs, or imagine scenarios that have not appeared in the records. Ema says the first two methods use existing evidence, while imagined scenarios are more experimental and need a new test case before they can pass evaluation.
Ema says proposals are stored with their evidence and evaluation results, and do not edit an AI Employee directly. An administrator can review and approve a proposal, apply it as a live trial or directly, and roll back an applied change. These are Ema’s descriptions of its design, not independently reported results from customer deployments.
Why it matters
Workplace agents can affect records, inboxes and decisions that cannot simply be reset like a coding puzzle. Ema’s approach tries to make improvement safer by separating learning from action: review what happened, test a proposed change against evidence, then leave the decision to an administrator.
That also offers a practical answer to a familiar problem with enterprise AI: useful lessons can be scattered across systems and lost when a task ends. Ema says its Enterprise Context Graph connects those records over time. Whether that produces better work depends on the quality of the feedback and tests, not the sleep metaphor, charming though it is.
Our read
The strongest part of Ema’s proposal is the boundary between suggesting and doing. Replaying real inputs or checking recorded decisions is more grounded than asking an agent to imagine its way to a better policy; imagined scenarios are explicitly the more experimental route.
For organisations considering workplace agents, the useful question is whether the evidence trail makes each proposed change understandable and reversible. Ema has described the machinery, but the post does not establish how well it performs in real deployments.
What to watch
- Whether customers report measurable improvements from changes proposed through Dream RSI.
- How administrators review evidence and choose between a live trial and direct application.
- Whether imagined scenarios produce useful test cases without creating false confidence. Activity teaser: Ema says its workplace AI agents can learn from past decisions by proposing changes for human review, rather than editing live workflows on their own. The system it describes checks proposals against recorded outcomes, replays past requests or tests imagined scenarios, with the last method treated as more experimental. It is a thoughtful attempt to give agents a learning loop without handing them the keys to the filing cabinet. The full story explains what Ema says is tested, what stays under human approval and what still needs evidence from real deployments.
Discussion spark: For workplace AI that can change real records, should administrators approve every improvement, or should systems be allowed to apply changes once they pass tests on recorded work?
Sources and evidence
- Dream RSI: AI Employees That Learn in Their Sleep | Ema (19 November 2026, 16:30 UTC)
Independent WittyWires tracker for public updates about Ema. Not affiliated with or endorsed by Ema; this is not an official account.