Salesforce argues the next leap in enterprise agents is automated, governed self-improvement

Salesforce is using its latest company essay to make a specific bet about where enterprise AI agents will separate from one another over the next few years. The company argues that the winning agents will not simply be those built on the newest foundation models. They will be systems that can learn from their own performance, through a controlled feedback loop that keeps improving outcomes over time.

The core idea is a contrast between two teams that ship agents on the same schedule with the same model and the same initial accuracy. Three months later, Salesforce says one agent is “dramatically better and noticeably cheaper to run,” while the other needs slow, manual patching to keep up with changing user needs, updated model capabilities, and evolving expectations.

Salesforce’s argument is that the difference is not the foundation model itself. It claims both teams receive a model upgrade, for free, and yet the gap persists. The only change, in the essay’s framing, is how quickly the first agent learns from user interactions and improves how it addresses those needs.

In current enterprise deployments, Salesforce says, agents are often treated like static applications. When performance degrades or recurring failure patterns emerge, the fixes typically come from human subject matter experts who diagnose issues and then patch the system, with rollbacks if changes fail. That approach, Salesforce argues, cannot scale across the volume of tasks and contexts that enterprise agents face.

To replace that manual cycle, the company describes recursive self-improvement as an automated process that detects what is failing, diagnoses root causes, tests multiple improvements using simulations, and retains the changes that improve both technical and business key performance indicators. Salesforce says this matters because it cannot “learn on its own” from usage volume alone, and no team can practically tune every recurring failure by hand at reasonable cost.

Salesforce also ties the thesis to its own deployment footprint. The company says more than 11 million Agentforce calls run each day, with no two sessions the same. It argues that the “experience” that can compound is the enterprise’s own production traffic, not the rented model weights, which the essay describes as depreciating quickly as frontier models improve and compute becomes cheaper.

A major practical point in the essay is that learning cannot be just “more autonomy.” Salesforce emphasizes that enterprises must define what “better” means before an optimization loop runs, including success metrics such as accuracy, speed, and cost and business KPIs, plus guardrails covering policy violations and regressions. The company frames the ability for owners to inspect or revise those definitions as a business and product decision rather than a purely technical one.

Salesforce also outlines what it calls an optimization engine that searches over agent designs without changing the frontier model weights themselves. It describes iterative evaluation where candidate changes are tested in simulations, then promoted only if they improve outcomes without violating constraints. The essay says the changes tend to be “larger than a prompt but smaller than an entire product,” focusing on the system around the model, including prompts, tool configurations, knowledge retrieval strategies, evaluators, and permission structures.

Safety concerns are another theme. Salesforce warns that an automated loop can learn the wrong thing if evaluators measure the wrong target, a scenario the company labels “reward hacking.” It also points to a separate risk in which recursive training on generated data can degrade behavior. Salesforce’s recommended structural safeguard is to verify proposed changes using external environments and multiple forms of evidence, including simulations, regression suites, adversarial cases, and often human judgment, while also monitoring evaluator calibration and drift.

Salesforce closes by saying that when weights are frozen, agents can still improve substantially because much of the adjustable system is outside the model itself. The company cites its AI research work as an early demonstration of reinforcement-learning-like techniques for optimizing a frozen-weight agent, referencing a 2023 Retroformer model that tuned prompts in response to new environments without updating weights. The essay suggests that enterprises should watch for systems that can compound safely over time, with gains recorded and auditable, rather than simply chasing each new foundation model release.