Using Full agent distillation in the platform dashboard? Start with
Train your complete agent for the current setup and
serving requirements. The golden-run workflow described below remains separate.
The problem
Production agents run the biggest model on every step because nothing smaller can be trusted with the workflow. Most of what a mature agent does is deeply repetitive: the same tools, the same verify scripts, the same repair moves against the same gate failures. You pay frontier prices for behavior a task-specific model can learn from a few dozen recorded sessions of that agent doing that job.How it works
1
Route traffic through ReasonBlocks
Set the gateway URL we issued you before running your agent:With Two lines, and nothing else in your agent code changes. This is what makes a run
provably complete later, and what most of your training data depends on — see
Integrating your agent.
ANTHROPIC_BASE_URL pointed at your hosted ReasonBlocks gateway, your agent speaks
its native protocol directly to ReasonBlocks. Your Anthropic API key passes through
untouched and is never stored. Every exchange is captured to hosted storage. The gateway
is byte-transparent and fail-open: if capture ever breaks, your traffic still flows.OpenAI, Gemini, AWS Bedrock, and other OpenAI-compatible hosts work the same way through
the same gateway with their own prefix — see Endpoint compatibility
for the full list and what each one gets you.Add the client shim in the same step, before your agent’s first LLM call:2
Traces reconstruct themselves
ReasonBlocks reconstructs raw traffic into verified runs on hosted storage: ordered steps,
tool calls paired to their results, and injected corrections identified structurally. Runs
are labeled golden (proven complete, no violations) or not. Only golden runs from
your workflow are eligible for training, and none of it requires a local database.
3
A cheaper task-specific model is trained
Once there are enough golden traces from this task, the hosted pipeline mines
(state → action) pairs from the complete runs and fine-tunes a smaller model for your
tenant on your workflow. Training is opt-in and tenant-scoped.
4
Your trained model goes live
The gateway serves your trained model through the same URL, with your frontier model
connected as backup. Which one answers is not a fixed confidence cutoff: a classifier
trained on your tenant’s own data scores the run and escalates to the frontier model
before a consequential action when your model’s answer should not be trusted. Escalated
episodes become training signal for the next cycle, so coverage widens over time. Same
ANTHROPIC_BASE_URL; your agent code unchanged. Your trained model is exclusively
yours — never shared with, or trained into, any other customer’s model.What makes it trustworthy
- Completeness is proven, not assumed. A run with a hole in it is never mined. The reconstruction layer proves each run was captured whole before the pipeline may quantify over it.
- Actions are validated against your workflow. Sampled steps must use known tools, real helper scripts, and well-formed arguments from vocabulary mined from your own traces.
- Whole-action selection only. Rejection sampling keeps or drops complete actions. Nothing selects on partial steps, which avoids low-quality training pairs.
- Escalation is a trained decision, not a guess. Whether to trust your model’s answer or escalate to your frontier model is decided by a classifier trained on your tenant’s own agreement data — not a fixed confidence threshold.
Get started
Quickstart
Gateway URL → env var → shim → capture → distill.
Integrating your agent
The gateway env var plus the two-line shim.
Endpoint compatibility
Anthropic, OpenAI, Gemini, and other LLM API paths.
Deployment
Hosted gateway, capture storage, and the distillation pipeline.