Today, we are introducing a new vision for LayerLens, a redesigned website, and Environments: synthetic replicas of production systems where agents can be evaluated and improved before they touch real ones.
Our vision centers on a new category: the Agent Readiness Platform. It brings together Agent Evals, Environments, and Synthetic Data around one mission: make AI agents ready for the work that matters.
The distinction is important. An intelligent model is not necessarily a production-ready agent. Knowing how a business works and being ready to work inside it are different capabilities. We are building the infrastructure to close that gap.
From answers to actions
Consider a support agent handling a refund. It explains the policy perfectly and writes a reassuring response. But it credits the wrong account, skips an approval, or leaves the invoice unchanged. The language is correct. The business state is wrong.
This is why evaluation must evolve. Benchmarks provide standardized comparisons of capabilities. Application-specific evals connect those capabilities to particular requirements. Agentic environments extend evaluation to something larger: sequences of actions, their consequences, and the resulting state of connected systems.
These approaches are complementary. What changes is the unit of measurement. We move from judging an answer to examining an execution trajectory.
Our conviction is straightforward: you cannot fully evaluate an agent’s ability to act without an environment in which its actions have consequences. A static test can examine its understanding of a refund. An environment can reveal whether it actually completed one.
Evaluation as AI’s fourth pillar
We believe evaluations should stand alongside compute, data, and models as the fourth pillar of AI.
Compute provides capacity. Data supplies examples and context. Models provide learned capabilities. Evaluation tells us whether those capabilities produce the outcomes we intended, and whether a change represents progress.
Without that feedback, scaling can mean doing the wrong thing more efficiently. Evaluation should therefore be foundational infrastructure, not the inspection station at the end of development. Agent Readiness makes that infrastructure specific to a task, a configuration, and the conditions of a business.
Evaluate. Simulate. Generate.
The LayerLens Agent Readiness Platform organizes this vision around three connected pillars.
Agent Evals measure behavior and outcomes. Observability captures execution traces, tool calls, retries, and errors. Deterministic scorers check measurable results; judges assess contextual requirements against rubrics. Together, they help teams understand not just whether an agent failed, but where its behavior diverged.
Environments recreate the operating conditions. They provide synthetic versions of supported systems in which agents execute tasks and change state. Fidelity is not cosmetic resemblance. It is whether interactions and state transitions behave correctly for the work being evaluated.
Synthetic Data supplies realistic records and inputs. It populates Environments and expands evaluation cases, connecting each Environment’s structure to situations an agent must handle.
Each pillar is useful independently. Together, they form an improvement loop: generate relevant inputs, execute inside an Environment, evaluate the result, and use the evidence to guide the next change. The output is more than a score. It is a direction for engineering.
Three ways to build an Environment
A generic application is not the same as your application. Your schemas, custom objects, approval rules, and workflows determine what an agent encounters. LayerLens addresses this with three types of Environments that share one runtime but draw on different sources of evidence.
Baseline Environments begin with public API specifications. LayerLens generates candidate tool implementations, executes them, and verifies changes between the initial and resulting state against expected behavior. Failed checks guide repairs; unresolved gaps remain explicit. This is the product as documented, without requiring access to a customer’s system.
Grounded Environments add the organization’s configuration through read-only metadata: schemas, custom objects, validation rules, and approval structures. Rather than copying production records, the architecture generates the differences from the public baseline. Verification compares supported responses against the real instance. This is the product as configured for your business.
Behavioral Environments incorporate recorded execution traces. They reconstruct observed interactions through replay, entity recall, and evidence-backed response templates. When the available evidence cannot support a response, the simulator abstains rather than inventing behavior. Held-out calls test how faithfully it reproduces those interactions. This is the business as it operates.
The common principle is measured fidelity. An agent’s score means little if the Environment rewards behavior that would fail in production. The Environment itself must earn trust.
From an evaluation to an improvement engine
Inside an Environment, teams can hill-climb against task-specific evals. Define success, run an agent, inspect its failures, change the relevant component, and run again. Improvements might involve prompts, tools, workflow logic, model selection, or supported training methods. Execution alone does not update the agent.
For the refund example, success is not a convincing message. It is the correct account credited, the invoice updated, and the required approval respected. Fresh validation cases then test whether the improvement extends beyond familiar examples.
The architecture also connects simulation to production: verifiers developed inside an Environment can assess production traces against the same criteria. Readiness becomes something teams can reassess as agents and systems change, rather than a label attached at launch.
Your business, not somebody else’s leaderboard
This release includes more than 15 new Environments spanning popular SaaS systems, databases, and consumer products. Companies can also create their own Environments around existing systems, MCP interfaces, and APIs. The catalog is a starting point, not the boundary of the platform.
We are already working with frontier AI labs, agentic companies, and large enterprises evaluating their agents in LayerLens Environments.
We see a future in which evaluation environments evolve with the businesses they represent: new tools expand the Environment, observed failures expand the tests, and every agent revision is measured against the work it must perform.
Our ambition is to make every important business workflow executable as an evaluation: a place to establish readiness, discover failures, and engineer improvement. The next benchmark for your agent should not be somebody else’s leaderboard.
It should be your business.