Resettable enterprise replicas for training and evaluating agents
Teams improving agents need realistic, stateful copies of business systems that reset after every run, and that supply is becoming a category of its own.
Crowded
Written for: Investor or analyst
Next step
Thesis: realistic, resettable environments become the scarce input for agent RL, much as labelled data was for supervised learning. The confirming signal would be frontier labs or software incumbents publicly licensing environment replicas, or a priced environment-as-a-service round. Already building: Microsoft with Agent Lightning, Kiln with Seahaven, the Trace2Env authors, and Arena on the evaluation side.
Why now
The training method is no longer the bottleneck. A builder working on the problem for more than a decade says the hard part is the evaluation environment, because shared production and staging systems cannot be reset. Arena raising $200M at a $3.1B valuation shows investors will pay for measurement. A statistical critique of the METR time-horizon plot shows that public benchmarks alone are not good enough.
What became possible
Microsoft's Agent Lightning v1.0 hooks existing agents into reinforcement-learning (RL) training without rebuilding them. Trace2Env uses a world-model agent to stand in for systems that cannot be reproduced, simulating them from logs of past activity. The Seahaven framework from Kiln provides realistic data, stateful writes and a reset to the same starting point on every run. MiMo-V2.6 shows frontier labs putting more compute into RL.
Two ways in
- System-of-record simulators
Resettable replicas of ServiceNow-, Salesforce- or SAP-style environments with synthetic but realistic data, used for agent RL and regression testing.
- Customer
- Teams building enterprise agents and frontier-lab data-buying teams
- Model
- Per-environment licence plus per-run compute fees
- Trace-to-environment service
Turns a customer's own agent logs into a simulated environment, in the style of Trace2Env, for private evaluation.
- Customer
- Regulated enterprises that cannot expose production to agent training
- Model
- Professional services setup plus a subscription
Competition
The trace-to-environment service is less taken. I found no company named Trace2Env and no product that compiles a customer's agent logs into a stateful, resettable simulator. The closest things are a USPTO patent filing on training surrogate tool models from captured traces, trace replay and trace-derived datasets in Braintrust and LangSmith, and Veris's VPC and on-prem mocked-tool platform. The opening is a product that runs fully inside the regulated customer's boundary. It would infer schemas and state transitions from logs, generate synthetic data that keeps privacy intact, and output a resettable environment with verifiers. The risk is that observability vendors or Veris and Collinear add this as a feature, so speed and compliance credentials matter more than the core technology.
Full competition notes (11 competitors)
The system-of-record simulator idea is crowded. Funded startups such as Arga, Fleet, Collinear, Veris and Halluminate, plus big data vendors like Turing, Centific and Mercor (through its reported Deeptune acquisition), already sell resettable Salesforce-, Workday-, Jira- and ServiceNow-style replicas to frontier labs and enterprises. ServiceNow's open WorkArena benchmark covers the basic ServiceNow evaluation case for free. A new horizontal replica vendor would struggle against them. A better angle is depth in one area they cover thinly: SAP or other ERP and finance back-office systems, regulated vertical systems like EHRs, core banking or insurance claims, or very high-fidelity business logic such as SAP posting rules that generic clones get wrong. The trace-to-environment service is less taken. I found no company named Trace2Env and no product that compiles a customer's agent logs into a stateful, resettable simulator. The closest things are a USPTO patent filing on training surrogate tool models from captured traces, trace replay and trace-derived datasets in Braintrust and LangSmith, and Veris's VPC and on-prem mocked-tool platform. The opening is a product that runs fully inside the regulated customer's boundary. It would infer schemas and state transitions from logs, generate synthetic data that keeps privacy intact, and output a resettable environment with verifiers. The risk is that observability vendors or Veris and Collinear add this as a feature, so speed and compliance credentials matter more than the core technology.
- Arga Labs
Raised a $10M seed led by General Catalyst in August 2026. Builds full digital twins of enterprise software such as Salesforce, Workday and email clients, with permissions and webhooks intact, and makes them easy to reset for RL. This is almost exactly the system-of-record simulator idea.
- Fleet AI
Builds RL gyms that replicate enterprise software such as Salesforce and Excel for frontier labs and enterprises. A third-party listing gives it about $15M in funding.
- Deeptune (acquired by Mercor)
Raised a $43M Series A led by a16z in March 2026 for high-fidelity training gyms covering accountant, support and DevOps workflows on tools like Slack and Salesforce. A directory listing and a Digg post say Mercor acquired it in July 2026, which would mean data vendors are buying into this space.
- Collinear AI
Runs simulation labs where LLM-played coworkers and users interact with simulated Jira, ServiceNow, Shopify, EMR and airline systems. A third-party profile says it works with ServiceNow and IBM, but I couldn't confirm that from a primary source.
- Veris AI
Raised an $8.5M seed for simulation environments with mocked enterprise tools. Enterprises use it to test agents and train them with RL. RL List says it can be deployed in the customer's VPC or on-prem, which overlaps with the regulated-enterprise private evaluation pitch.
- Halluminate
YC S25 company whose Westworld product is a simulated internet with Salesforce-, Slack- and ticketing-style apps and verifiable RL rewards. One search summary says it now positions itself around financial services.
- Turing RL Gyms
A large data vendor selling interactive replicas of Jira, Salesforce, Zendesk, HubSpot and other apps. It ships them in Docker containers with step-level verifiers, mainly to frontier labs.
- Centific RL Environments-as-a-Service
Sells enterprise RL environments covering D365 CRM, Outlook, Teams, SharePoint and Workday.
- ServiceNow WorkArena / BrowserGym
An open-source ServiceNow-based benchmark of thousands of knowledge-work task instances, plus the WorkArena++ follow-up. It is a free baseline for ServiceNow-style evaluation.
- Maxim AI
Agent simulation and evaluation that can be deployed in the customer's VPC. It simulates users and personas rather than full stateful system replicas, but it competes for the private-eval budget.
- Braintrust / LangSmith
Observability platforms that already turn production traces into eval datasets and replay recorded traces. That is a lighter version of trace-to-environment, and it reaches the customer's existing logs first.
Checked 2026-10-09
Risks
- Frontier labs build environments in-house and buy only commodity data
- Simulated environments drift from real system behaviour
- Software incumbents may restrict replicas of their schemas or terms
Strong team fit; tech readiness and open space and buildability are the weak point (6).
tech 6 · pull 7 · space 6 · fit 8 · build 6
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
- Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real HarnessesMicrosoft Research
- From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment SimulationHF Daily Papers
- I built an open source framework for building RL environments. Named "Seahaven" after the fake town in The Truman Show.r/LocalLLaMA
- MiMo-V2.6: Scaling Reinforcement Learning Towards Self-ImprovementHF Daily Papers
- Popular AI leaderboard Arena nearly doubles valuation to $3.1B valuation in 10 monthsTechCrunch AI
- On the estimation and validity of AI time horizons---a statistical look at the METR plotarXiv
What happened next
- +7d · 16 Oct 2026 nothing yet. No launches, funding rounds, acquisitions, or significant press coverage of enterprise replica training environments for agents were found after October 9, 2026. The most recent developments in this space—including Patronus AI's Series B, Arga Labs's seed funding, and Centific's platform launch—all predate the flag.
Status history
- 2026-10-09 Flagged as buildable