Quickstart

For Full agent distillation in the platform dashboard, follow Train your complete agent. It uses an isolated tool environment, on-policy token scores and whole-task evaluation. This page describes the older hosted gateway and golden-run workflow.

ReasonBlocks hosts the capture gateway, storage, trace reconstruction, the distillation pipeline, and trained model serving. Your integration is one environment variable and two lines of code — nothing runs on your own servers.

1. Get your gateway URL

We issue your tenant gateway URL when you onboard — reach out if you do not have one yet. It looks like:

export ANTHROPIC_BASE_URL="https://gateway.<your-tenant>.reasonblocks.com"

Use the exact URL shown for your tenant. Do not substitute localhost.

2. Set the environment variable and run your agent

Set ANTHROPIC_BASE_URL before starting your agent. Use your existing Anthropic API key as you always have. ReasonBlocks passes it through to upstream and never stores it.

export ANTHROPIC_BASE_URL="https://gateway.<your-tenant>.reasonblocks.com"
export ANTHROPIC_API_KEY="sk-ant-..."   # your existing key; unchanged
python your_agent.py                     # nothing inside your SDK or framework changes

Your agent speaks the Anthropic Messages protocol to ReasonBlocks when using ANTHROPIC_BASE_URL. OpenAI and other providers use the same gateway with different base URL env vars. See Endpoint compatibility. The gateway is byte-transparent and fail-open: if capture ever breaks, your traffic still flows. This step is half the integration. The other half is the two-line client shim in step 3, which labels run boundaries. Both are required: without the shim the gateway still captures traffic, but run boundaries are inferred rather than proven and far fewer runs qualify for training.

3. Add the client shim

Run boundaries are what make a trace minable, and the shim is what states them. Install it in your agent environment:

pip install rbtrace

Then follow Integrating your agent for rbtrace.client.install(), plus with rbtrace.client.run(): per job when one process handles many. The gateway URL and capture path are unchanged.

4. Training

Training is opt-in and tenant-scoped. Once there are enough golden runs for this workflow, ReasonBlocks mines policies from them and fine-tunes a cheaper task-specific model on your workflow, within a spend ceiling. When the job finishes, you get a report with the model identifier and whether it’s ready to go live.

5. Go live with your trained model (same env var)

Keep ANTHROPIC_BASE_URL pointed at the same gateway URL — nothing in your agent changes when your trained model starts answering. Per request, a classifier trained on your tenant’s own data decides whether your model’s answer can be trusted or whether the gateway should send that request to your frontier model instead. Escalated requests are captured as edge cases and feed the next training cycle, so coverage grows from the runs you keep capturing. The agent still speaks Anthropic Messages. ReasonBlocks serves the flat transcript contract from the training report.

Getting Started

Reference

1. Get your gateway URL

2. Set the environment variable and run your agent

3. Add the client shim

4. Training

5. Go live with your trained model (same env var)

Next steps

Integrating your agent

Deployment

​1. Get your gateway URL

​2. Set the environment variable and run your agent

​3. Add the client shim

​4. Training

​5. Go live with your trained model (same env var)

​Next steps

Integrating your agent

Deployment

1. Get your gateway URL

2. Set the environment variable and run your agent

3. Add the client shim

4. Training

5. Go live with your trained model (same env var)

Next steps