Fleet is a platform from US-based Fleet AI, Inc. for training and evaluating AI agents. It provides simulation environments that faithfully reproduce the interfaces of business software such as Salesforce and Excel, and has agents perform work-like tasks inside them in order to measure their behavior and improve it through reinforcement learning. The official site describes its work as using “simulated worlds and real-world challenges to better understand and shape the behavior of artificial intelligences.” Its main customers are frontier AI research labs and large enterprises in sectors such as finance and insurance. It is not an app for everyday users, but infrastructure for teams that build agents.
Key Features
- Learning environments that reproduce business software: Fleet provides environments that simulate real software operations, including CRMs, spreadsheets, and browser or desktop workflows. You can test how an agent operates without letting it touch production systems
- Turning software into reinforcement learning environments: Fleet has a mechanism for converting arbitrary software into a gym-style interface usable for reinforcement learning (RL). On GitHub it also publishes a fork of the related project “gym-anything”
- Python SDK and API: Environments can be created and managed, and their state retrieved, from code. Note that the Python SDK repository (fleet-ai/fleet-sdk, Apache 2.0) had been archived (made read-only) as of August 2026, so check the official documentation for the current way it is distributed
- Evaluation framework “Harbor”: Fleet publishes on GitHub a fork of Harbor (Apache 2.0), a framework for running agent evaluations and building and using RL environments, and is also developing “harbor-train,” a training layer that combines SkyRL and SkyPilot
- Industry-specific custom environments: For sectors such as finance and insurance, Fleet works with customers to build dedicated evaluation and training environments tailored to each industry’s workflows and rules
Pricing
| Plan | Price | Notes |
|---|---|---|
| Public plans | None | No pricing page was found on the official site |
| Adoption | Check the official site | Appears to assume individual arrangements with enterprises and research labs |
Pricing information is as of September 2026. No published pricing plans could be confirmed. For the latest information, please check the official website.
Pros & Cons
✅ Pros
- You can verify how well an agent handles real screen operations without putting production business systems at risk
- Because agents can be trained in environments close to your own business software, it is easier to judge whether they will be useful in practice, rather than relying on scores from general-purpose benchmarks
- The evaluation framework is open source, making it easier to integrate into an in-house evaluation stack or to check its behavior locally
- Even in heavily regulated, rule-bound industries such as finance and insurance, it may be possible to prepare environments that reflect those constraints
⚠️ Cons
- Pricing is not published, so you need to contact the company to get a sense of cost
- Customers are mainly research labs and large enterprises, and there is no visible sign-up path that lets individual developers or small teams try it right away
- The official documentation requires a login, which makes it hard to study the details before signing a contract
- With the Python SDK repository archived, it is harder to follow development from the outside
- Many other services share the name “Fleet” (JetBrains Fleet, Fleet Device Management, and others), so it is easy to mix them up when searching for information
Comparison with Similar Services
| Service | How it differs from Fleet |
|---|---|
| Mechanize | Focuses on high-fidelity environments for coding work. Fleet centers on reproducing business software such as CRMs and spreadsheets |
| Surge AI | Covers a broad range of data labeling and human feedback (RLHF). Fleet specializes in building environments for operating business software |
| Bespoke Labs | Centers on open-source tools for evaluation and data curation. Fleet’s emphasis is on building and providing the environments themselves |
| ServiceNow (WorkArena) | Centers on agent evaluation benchmarks built around its own products. Fleet provides third parties with environments for a wide range of business software |
Who Is It For
- AI research labs that want to secure business-software environments and data for post-training LLMs and VLMs
- Enterprise development teams that want to measure accuracy in a safe place before handing their CRM, ERP, or spreadsheet operations over to AI agents
- Enterprises that need evaluation environments aligned with industry-specific rules, such as in finance and insurance
- Engineers who want to learn how agent evaluation and RL environments are built by reading open-source code
Summary
When you want AI agents to operate business software, the biggest hurdle is this: testing in production is risky, but general-purpose benchmark results do not tell you how well an agent will actually perform the work. Fleet fills that gap by providing a practice ground that reproduces real operating screens. On the other hand, pricing and sign-up methods are not public, and the documentation requires a login, so adoption starts with contacting the company. It is an option for organizations seriously planning to bring agents into their operations, rather than for individuals who want to try it out.
Related Links
- Official website
- Documentation (login required)
- Official blog (Notes)
- GitHub
- X (formerly Twitter)