Train yoursupport agentbefore you deploy. Keep training it after.

No. The core loop of scenario generation, simulation, root-cause analysis, and regression gating works with any model behind any API, closed or open. The open-source requirement only applies to the last step: RL fine-tuning, where the model weights have to be yours to update. On any model, every run ends in an improvement report with proposed prompt and harness fixes: concrete edits to instructions and tool definitions you can ship directly. On open models, the same verified runs additionally become RL rewards.

Three sources: your code and prompts (read directly), your documents (PRDs, policies, regulations), and your production traces (pulled in through integrations with major observability platforms). Scenarios seeded from production traces are replays of things that actually happened to your agent, so realism is inherited rather than assumed. Every scenario ships with a verifier that defines what passing means, so you can audit the criteria before you trust the score.

Reliable enough to gate deploys on. Wherever an outcome can be checked objectively (a tool call made, a record written, a number matching), verifiers check environment state, not vibes. Where judgment is required, graders are framed around specific failure modes rather than open-ended quality, which in our published comparison reached 94% consistency versus 66% for conventional grading, and we run consistency tests on every grader before its verdicts count. The criteria are yours to read and edit, so you can audit what passing means before you trust the score.

An eval reports a score; the loop produces changes. When a run fails, Veris replays it, isolates the decision that went wrong, and points at the instruction or tool definition that caused it, with a proposed fix. The failure is then pinned as a permanent regression scenario, so the fix can't quietly rot. Evals tell you the agent is at 74%; the loop tells you why, what to change, and proves next week's version didn't reintroduce it.

A production incident lands in your observability stack as a trace. Veris imports the trace, reconstructs the environment state that produced it, and generates a scenario set around it: the original failure plus variations that stress the same weakness. Those scenarios join the regression suite that runs on every push. Median time from incident to pinned regression coverage is measured in minutes, not sprint cycles.

It plugs into it rather than replacing it. Connect the platform you already use (Datadog, an OpenTelemetry pipeline, or an LLM-observability tool) and production traces flow into Veris as raw material: incidents become replayable scenarios, and real usage patterns steer scenario generation. Your instrumentation stays exactly as it is; Veris consumes what it already emits.

It stays under your controls. Traces can be redacted before Veris sees them, scenario generation preserves the shape of an incident rather than the customer inside it, and the whole pipeline runs in your VPC or on-prem if that's where you deploy. Retention is configurable, and your traces are used for your scenarios and your training, nothing else.

Every simulated run is already a labeled training sample: the scenario's verifier graded it. When your model is open-source, Veris uses those verified outcomes as rewards for reinforcement fine-tuning (GRPO and similar), targeting exactly the tasks your agent keeps failing. Training runs in your infrastructure or ours, with the same VPC deployment options as the simulation engine, so weights and traces never leave your boundary unless you want them to.

No. That's what the regression suite is for. Improvement targets the failure modes, and every candidate, whether a tuned prompt or tuned weights, is evaluated on held-out scenarios plus the full regression suite before it ships. In our published incident case study, the improved agent fixed the failure while holding performance on every prior scenario, without overfitting to a single example. If a candidate regresses anywhere, the gate stays shut.

At any point in the agent's development. Before a line of agent code exists, scenarios generated from your docs define what it must handle. During development, the simulation loop is your QA gate. At launch, it's the evidence for review. In production, incidents feed back as regression scenarios and training signal. Teams enter the loop wherever they are, and the rest of it attaches naturally.

Repo access (or just the prompts and tool definitions), and optionally a connection to your observability stack for production traces. The first scenario suite is typically running within a day: connect the repo, generate scenarios, run the loop. No framework migration: Veris sits around your agent, not inside it.