llms.txt
Compare with other sites' llms.txt files →# Veris AI
> Veris AI is simulation infrastructure for enterprise AI agents: a high-fidelity, stateful twin of a company's systems, data, APIs and users, sealed off from production. The same infrastructure is used two ways — to build customized benchmarks that test the agents an enterprise buys and the agents it builds against its own operating conditions, and to hand coding agents a per-agent environment to verify their own work against before the PR.
**Veris is one piece of infrastructure — a simulated twin of an enterprise's systems and users — put to two distinct uses.** Which one applies depends on whose agent is being tested and when.
**Use 1 — Customized benchmarks for the agents an enterprise buys and builds.** Veris stands up a twin of the company's production estate, turns that company's own policies, controls and regulations into the conditions every candidate has to pass, and runs a vendor's AI and an in-house build through the identical test in the identical environment. This is a benchmark built around one enterprise's workflows and rules, not a generic public eval: vendor claims that cannot otherwise be checked become measurable, and the output is a decision packet (what each system did, where it failed, cost at real volume, a reproducible record) rather than a dashboard. Anything that makes a decision qualifies — a model, a copilot, a document reader, an agent, a rules engine with an LLM attached.
**Use 2 — A dependency environment coding agents test their own work against.** Veris gives each coding agent its own stateful simulation of the dependencies its code calls (databases, payment APIs, CRMs, queues, internal services, MCPs), spun up, mutated, reset and torn down inside the agent loop. It plays the role a staging environment would, except one per agent and disposable, so thousands run in parallel without contending over shared state. Static mocks return canned responses; a Veris dependency holds state, so writing a row and reading it back, or waiting on a webhook, is a real test. The agent catches its own integration bugs and opens a PR that already proved itself.
Both uses run on the same primitives:
- **Twin** — a stateful simulation of the systems an agent depends on (APIs, tools, databases, third-party services), not recorded fixtures or static mocks. Spun up, mutated, reset and torn down per run, and air-gapped from production by construction rather than by configuration.
- **Simulated user** — an LLM-driven counterpart on the other side of the conversation, used in place of scripted turns or human testers.
- **Scenario** — one task plus the world state it starts from, generated from source code, docs, policies or production traces rather than hand-written, each paired with a verifier that checks what the agent did to the environment, not just what it returned.
- **Public benchmarks:** Veris runs its own boards on this infrastructure and publishes them at veris.ai/leaderboard. The headline is the Veris FinAgent Arena, one index over AML Bench and TreasuryQA Bench for agents working in a simulated bank; beside it, VAmoS Pro Bench grades voice agents on a regulated utility's billing line. Both are listed under Public benchmarks below.
**Security and compliance.** Veris is SOC 2 Type II audited (AICPA) and GDPR compliant, and meets HIPAA requirements. Simulations run on synthetic data by default, so no production customer data is needed, and the platform deploys into your own VPC (AWS, GCP, Azure), on-prem, or Veris Cloud.
## Start here
- [Veris AI](https://veris.ai/): Overview of both uses: a simulated copy of your systems, data, APIs and users, where AI is tested for quality and compliance, agents are trained, and coding agents get a stack to build against — all before anything touches production.
- [Platform](https://veris.ai/platform): The shared machinery underneath both uses, end to end — simulated users, simulated systems, AI-composed scenarios, isolated sandboxes at scale, and root-cause analysis that turns a failure into a fix.
- [FAQs](https://veris.ai/faqs): Direct answers to the questions Veris gets asked: what it actually simulates, how a twin is isolated from real systems, where simulation data comes from, how it differs from traditional evaluation and from grading production traces, how it fits alongside an existing agent platform or observability stack, plus pricing model, ROI, security and compliance (SOC 2 Type II, GDPR, HIPAA) and who buys it.
## Public benchmarks (the Veris leaderboard)
- [Veris FinAgent Arena](https://veris.ai/leaderboard/arena): The headline board. One index, out of 1,000, over the two financial boards for the 10 model-and-harness pairs that ran both AML Bench and TreasuryQA Bench in the same simulated bank: a work score per bench from pass@1 and pass^3, combined and reduced by a cost penalty. Claude Opus 5.5 on Claude Code leads at 609. Full table, charts and a sourced methodology; the per-task detail stays on each bench.
- [VAmoS Pro Bench leaderboard](https://veris.ai/leaderboard/vamos-pro): The voice board Veris maintains: 16 voice agent stacks on a regulated utility's billing line, 100 chained billing and payment-assistance calls three times each over street, café and television noise, graded on the world each call leaves behind rather than the transcript. Phonic leads at 57.5% task completion and the field runs down to 17.3%. Phonic, added on 30 October 2026, is the newest stack. Recordings of a failed call per stack, a cost and completion frontier, and the exclusions audit.
- [AML Bench leaderboard](https://veris.ai/leaderboard/aml-bench): One of the Arena's two components: twelve agent candidates (harness plus pinned model) working 153 anti-money-laundering investigations three times each as a second-level investigator, against Apache Fineract, Marble and a sanctions watchlist sharing one frozen synthetic bank. Graded on the disposition filed in the case system, not on the chat. Carries the per-task table, a failure gallery and published methodology.
- [TreasuryQA Bench leaderboard](https://veris.ai/leaderboard/treasuryqa): The Arena's other component: ten agent candidates (model plus pinned command-line harness) answering 70 exact-figure questions over 1,435 image-only scans of U.S. Treasury publications, three times each, reached through a Dropbox twin's MCP server and read with a hosted OCR server. Graded by a script rather than a model: the figure the reply commits to must fall in a range built to exclude every known wrong answer. Carries a failure gallery, a scan-quality breakdown and published methodology.
- [Leaderboard](https://veris.ai/leaderboard): Overview of every public board with the top of each: the FinAgent Arena, VAmoS Pro Bench, AML Bench, TreasuryQA Bench and VAmoS Bench.
- [VAmoS Bench leaderboard (the earlier voice board)](https://veris.ai/leaderboard/vamos): The first VAmoS board, card support rather than utility billing: seventeen customer-facing voice agent stacks scored on end-to-end task completion, with audio samples and per-scenario breakdowns. Superseded by VAmoS Pro Bench for new results; its per-vendor framework reports are listed under Optional.
## Use 1 — Customized benchmarks for enterprises (agents you buy and agents you build)
- [Veris for benchmarking](https://veris.ai/veris-benchmark): The main page for this use. How a benchmark is built around your own policies, workflows and systems: simulated users, sandboxed twins of your services, scenario generation, automated judges that grade the state the systems are left in, and what an engagement looks like week by week.
- [Twin fidelity benchmark](https://veris.ai/fidelity): How closely the twins behave like the production systems they stand in for, scored by an independent instrument: 703 of 721 verdicts agree, every mismatch named.
- [Test and Compliance](https://veris.ai/solutions/test-and-compliance): The compliance angle on the same use: a vendor's AI and your own build facing the same conditions in an air-gapped twin of production, your policies and regulations turned into the test, and a report your risk function can read.
## Use 2 — A test environment for coding agents (agentic SDLC)
- [Agentic SDLC](https://veris.ai/solutions/agentic-sdlc): The main page for this use. A stateful dependency sandbox handed to each coding agent: every outbound call to a database, internal service or third-party API hits a live simulated instance instead of the real system, so the agent runs genuine integration tests and kills the bug before the PR rather than in review. Works with existing coding agents and existing compute sandboxes — it fills the sandbox rather than replacing it.
- [Coding agents, with and without Veris](https://veris.ai/side-by-side): The controlled benchmark behind that claim: same issue, same model, same repository, only dependency access changes.
## Compare, guides, security and licensing
- [How to test an AI agent before production](https://veris.ai/guides/testing-ai-agents-before-production): The five checks a pre-production test has to cover, the tools for each, and what a controlled sandbox with stateful twins, simulated users and injected faults adds that evals and tracing cannot.
- [The vendor's simulator vs an independent bench](https://veris.ai/compare/vendor-simulations): What Sierra Simulations, Google Agent Platform evaluations, Decagon Simulations and Agentforce Testing Center do well; the five things a vendor's own simulator cannot do by construction (one exam across vendors, pass criteria you sign, stateful twins behind the agent, evidence that outlives the contract, a champion-challenger review); how a bake-off runs; six questions to ask any vendor.
- [Veris with LangSmith, Braintrust, Arize, Langfuse and promptfoo](https://veris.ai/compare/observability-and-evals): What each tool is for, what none of them provides, a side-by-side table, and how they fit together with Veris in one loop. They trace and score; Veris is the environment.
- [How Veris compares](https://veris.ai/compare): Index of the comparison pages.
- [Security and deployment](https://veris.ai/security): SOC 2 Type II, HIPAA, GDPR; SaaS or fully on-prem inside your VPC on AWS, GCP or Azure; synthetic data by default; what a security review receives.
- [How Veris is licensed](https://veris.ai/pricing): Per bench or an annual platform licence, VPC deployment priced separately, what sets the price, and a first bench in about two weeks. No public price list: contact us for pricing.
## VAmoS Pro Bench reports
- [VAmoS Pro Bench: a harder exam for voice agents, and what fourteen stacks did with it](https://veris.ai/blog/vamos-pro-bench-a-harder-exam-for-voice-agents): Our new voice-agent benchmark puts fourteen stacks on a utility billing line: two to four requests per call, angry callers, a television in the room, and a verdict taken from the world the agent…
- [Your voice agent is talking to someone watching TV. How bad could it be?](https://veris.ai/blog/ai-coustics-voice-focus): With a TV playing behind the caller, our voice agent failed nearly every call.
- [A television in the room breaks most voice agents](https://veris.ai/blog/vamos-pro-findings-tv-in-the-room): On VAmoS Pro Bench, a television playing behind the caller cut pooled task completion from 38.7% to 8.6%.
- [Decision models for voice-agents: a benchmark](https://veris.ai/blog/decision-models-turn-detection): We put eight decision models in charge of when a voice agent is allowed to talk, across 2,694 simulated phone calls.
- [Q: Is Jev useful for voice agents? A: Let's benchmark it](https://veris.ai/blog/jev-turn-detection): Text agents never have to guess when the user has finished talking.
## Writing
- [Veris joins the Fireworks Frontier Ecosystem](https://veris.ai/blog/veris-fireworks-frontier-ecosystem): Fireworks trains and serves open models. Veris builds the world the model is tested and trained in. From an empty workspace to a model that climbs your own benchmark, in four steps.
- [Voice AI Agent Guardrails: How to Test and Tune Them in Simulation with Veris AI](https://veris.ai/blog/voice-ai-agent-guardrails-test-and-tune-in-simulation): We put a first-pass Amazon Bedrock Guardrails config on Rory, a voice billing agent, ran it through simulated adversarial and ordinary customer calls, and tuned it policy by policy.
- [Evals vs. Benchmarks: Every Agent Attempt Needs Its Own World](https://veris.ai/blog/evals-vs-benchmarks): Evals grade the answer.
- [Why Enterprises Need a Centralized AI Evaluation Platform](https://veris.ai/blog/centralized-ai-evaluation-platform): Compare AI vendors fairly, enforce company-wide standards, and test in-house agents with a centralized, vendor-neutral AI evaluation and benchmarking platform.
- [E2B + Veris: Give Coding Agents a Computer and a World to Work Against](https://veris.ai/blog/e2b-veris-a-computer-and-a-world-for-coding-agents): E2B gives every coding agent its own secure cloud computer.
- [Don't upgrade the model, simulate first](https://veris.ai/blog/dont-upgrade-the-model-simulate-first): A few prompt changes can beat a model tier upgrade at one-eighth the cost. But only if you have a robust way of measuring it. Here's how to find that out for your own agent before you pay for it.
- [Real People Don't Talk Like That: Veris AI vs τ²-bench on user-simulation realism](https://veris.ai/blog/user-simulation-realism-veris-vs-tau2-bench): Veris AI's simulated users score 84.8% on a five-check human-likeness rubric against τ²-bench-retail's 68.6%, a 16.3-point lead, ahead on all five checks and all three behavioural dimensions, judged…
- [The Right Model for the Right Task](https://veris.ai/blog/the-right-model-for-the-right-task): Six models, one production telecom assistant, three scenario sets. Three different winners, and the cheapest model in the pool won one of them outright.
- [The loop is easy, feeding it real failures is the hard part](https://veris.ai/blog/the-loop-is-the-easy-part): Building an agentic self-improvement loop is easy: run → diagnose → fix → repeat.
- [Let OpenClaw Run Wild in Simulation, Not on Your Customers](https://veris.ai/blog/let-openclaw-run-wild-in-simulation): Veris comes up with realistic scenarios you wouldn't, runs them in parallel, and surfaces the failures you don't know to look for.
- [Battle-Testing Enterprise AI Agents: Veris AI × AWS Bedrock AgentCore](https://veris.ai/blog/battle-testing-enterprise-ai-agents-veris-ai-aws-bedrock-agentcore): A practical guide to shipping enterprise agents that work: using AWS Bedrock AgentCore to build and deploy, and Veris AI to simulate, grade, and improve.
- [Veris Plus is live](https://veris.ai/blog/veris-plus-is-live): Our simulation sandbox that's been powering enterprise AI agents for a year is now fully self-serve.
- [RFT Can Make Agents with Small Models Outperform Big Models on Enterprise Tasks](https://veris.ai/blog/rft-can-make-agents-with-small-models-outperform-big-models-on-enterprise-tasks)
- [Veris AI Announces Availability on Google Cloud Marketplace to Deliver Reliable, Production-Ready AI Agents](https://veris.ai/blog/veris-ai-announces-availability-on-google-cloud-marketplace-to-deliver-reliable-production-ready-ai-agents)
- [Your LLM-as-a-Judge is Asking the Wrong Question](https://veris.ai/blog/llm-as-a-judge)
- [SantaBench: Agents Roasting on an Open Fire](https://veris.ai/blog/santabench): We built SantaBench, a fun benchmark with a serious methodology.
- [Never Waste a Good Failure: How Veris AI Turns Production Incidents into Self-Improving Agents](https://veris.ai/blog/never-waste-a-good-failure-how-veris-ai-turns-production-incidents-into-self-improving-agents): Turn production failures into improvements. See how Veris AI uses simulation, targeted evaluation, and automated prompt optimization to help agents self-improve without regressions.
- [Environments, Not Examples: How Reinforcement Fine-Tuning Trains Enterprise-Grade Domain-specific Agents](https://veris.ai/blog/environments-not-examples-how-reinforcement-fine-tuning-trains-enterprise-grade-domain-specific-agents): See how reinforcement fine-tuning transforms small models into high-accuracy Sigma rule generators—delivering enterprise-grade detection at lower cost and latency.
- [Technical Report: How Reinforcement Fine-tuning Trains Enterprise-Grade Domain-specific Agents](https://veris.ai/blog/technical-report-reinforcement-learning-fine-tuning-for-enterprise-ai-agents): Deep dive into how Veris AI uses reinforcement fine-tuning and simulation to train small models to generate accurate Sigma detection rules for enterprise security.
- [Dynamic Benchmarking: Evaluate AI Agents through Environments, not Datasets](https://veris.ai/blog/dynamic-benchmarking)
- [Building an Agent? You Need an Environment](https://veris.ai/blog/building-an-agent-you-need-an-environment): Agents live in messy systems with unpredictable users, interconnected tools, and cascading consequences.
- [Introducing Veris AI: A New Way to Train Enterprise AI Agents Through Simulated Experience](https://veris.ai/blog/introducing-veris-ai-a-new-way-to-train-enterprise-ai-agents-through-simulated-experience): We're thrilled to share that Veris AI is launching with $8.5M in seed funding from an oversubscribed round led by Decibel Ventures and Acrew Capital.
- [BankingBench: Measuring Model Performance](https://veris.ai/blog/bankingbench-measuring-model-performance)
## Optional
- [Company](https://veris.ai/company): Who Veris is, and investors.
- [Book a demo](https://veris.ai/demo)
- [Support](https://veris.ai/support)
- [Privacy policy](https://veris.ai/privacy-policy)
- [Terms and conditions](https://veris.ai/terms-and-conditions)
- [VAmoS Bench (earlier board): Deepgram on Veris AI's VAmoS Bench: one websocket for the whole stack, and a refusal that named the cardholder](https://veris.ai/blog/vamos-bench-deepgram): Deepgram's Voice Agent API completes 64.0% of tasks on VAmoS Bench, tenth of the seventeen voice stacks we tested, answers in 2.06 seconds and costs about $0.112 a call with speech recognition, model…
- [VAmoS Bench (earlier board): LiveKit on Veris AI's VAmoS Bench: 0.7 points behind first, on identical parts](https://veris.ai/blog/vamos-bench-livekit): LiveKit completes 70.3% of tasks on VAmoS Bench, second of the seventeen voice stacks we tested and 0.7 points behind pipecat on the same speech recognition, model and voice.
- [VAmoS Bench (earlier board): OpenAI Realtime on Veris AI's VAmoS Bench: the best first impression on the board](https://veris.ai/blog/vamos-bench-openai-realtime): OpenAI Realtime completes 81.1% of simple flows, the best of the seventeen voice stacks on VAmoS Bench, answers in 1.53 seconds and connected all 300 calls.
- [VAmoS Bench (earlier board): Gemini 3.1 Live on Veris AI's VAmoS Bench: the cheapest stack on the board, and the best at saying no](https://veris.ai/blog/vamos-bench-gemini-3-1-live): Gemini 3.1 Live costs about $0.016 a call, the cheapest of the seventeen voice stacks on VAmoS Bench, and answers in 1.38 seconds, second fastest.
- [VAmoS Bench (earlier board): pipecat on Veris AI's VAmoS Bench: the highest score on the board, assembled from three vendors](https://veris.ai/blog/vamos-bench-pipecat): pipecat completes 71.0% of tasks on VAmoS Bench, the best of the seventeen voice stacks we tested, at about $0.045 per call and with all 300 calls connected.
- [VAmoS Bench (earlier board): Your voice agent eval is grading the transcript, not the outcome](https://veris.ai/blog/voice-agent-eval-grades-the-transcript): The practical guides to voice agent testing agree on latency, ASR quality and interruption handling.
- [VAmoS Bench (earlier board): Vapi on Veris AI's VAmoS Bench: strong scores, and a call it should not have completed](https://veris.ai/blog/vamos-bench-vapi): Vapi completes 69.3% of tasks on VAmoS Bench, fourth of seventeen and above the field avera