Foundry

What is Foundry?

Foundry is a simulation and data platform for teams developing AI agents that operate websites. Browser-agent evaluation is difficult when live sites change, sessions cannot be reset, anti-bot systems interfere, and failures are only observed as a final pass or fail. Foundry addresses this with reproducible browser environments and structured records of agent behavior.

The platform combines simulation, evaluation, custom long-horizon datasets, and reinforcement-learning infrastructure. Its Agent Web Engine can provide a task, reset an environment, expose a browser connection, record events and state mutations, and compare the final state with known ground truth.

How to Use Foundry

1. Apply for access to the private beta and define the target browser-agent tasks.
2. Select or create reproducible environments that represent the relevant enterprise workflows.
3. Integrate an existing agent through Foundry's Python SDK and browser connection.
4. Reset an environment, run the task, and record events, state changes, failures, and the final result.
5. Score performance against ground truth and classify issues such as bad clicks, layout shifts, and misfires.
6. Use the evaluation data for regression testing, supervised fine-tuning, or reinforcement learning.

Core Features

• Reproducible simulation: Provides resettable browser environments intended to avoid uncontrolled web drift.
• Action-level evaluation: Tracks and classifies each agent action instead of exposing only a final outcome.
• State inspection: Records events, final state, and mutations for debugging and scoring.
• Custom datasets: Uses expert annotators to create long-horizon enterprise browser tasks.
• Reinforcement learning: Supports repeated trajectories in controlled environments without live-site anti-bot constraints.
• Python SDK: Connects Foundry environments to an existing browser-agent workflow.
• Ground-truth scoring: Compares the agent's final state with an expected result.
• Public benchmark: Provides benchmark results for comparing browser-agent performance.

Use Cases

• Benchmark browser agents on stable, repeatable enterprise tasks.
• Diagnose click, navigation, form, layout, and state-management failures.
• Generate supervised fine-tuning data for long-horizon web workflows.
• Train agents with reinforcement learning across many controlled trajectories.
• Run regression tests before releasing a new model, prompt, policy, or toolchain.

Pricing

Foundry is currently presented as a private beta and does not publish standard pricing. Teams must apply for access and discuss dataset, simulation, evaluation, and training requirements directly.

Frequently Asked Questions

Is Foundry a general website automation agent?

No. It is infrastructure for producing data, simulating environments, evaluating behavior, and training browser agents.

Why not evaluate only on live websites?

Live sites can change, throttle traffic, block automation, or produce inconsistent states. Reproducible simulations make comparisons and regression tests more reliable.

Does simulation prove an agent will work safely in production?

No. Production sites, permissions, user data, and failure modes can differ. Teams still need controlled rollout, security review, live monitoring, and human fallback.