FRONTIER LLM  //  AGENTIC RAILS

Rails you can trust in production.

Guardrails, orchestration, and observability for agents built on EchoNest. Define what's allowed, see every step, and stay in control when models act on their own.

Get API access Read the docs →
POLICY · checkout-agentv4
read customer recordallow
draft refundallow
issue refund > $500approval
export PIIblock
THE RAILS

Three layers between a model and your production systems.

01 · GUARDRAILS
Define what's allowed
Policy checks on every tool call and output. Block, redact, or require human approval before anything leaves your boundary.
02 · ORCHESTRATION
Stay in control of flow
Plan, route, and retry across models and tools — deterministic control flow wrapped around non-deterministic models.
03 · OBSERVABILITY
See every step
Every step, token, and tool call traced. Replay any run and debug failures long after they happen.
FULL-RUN OBSERVABILITY

Every run, on the record.

When an agent does something unexpected, you don't guess. Open the trace, see the plan, the policy decisions, and every tool call with timing — then replay it.

RUN TRACE · run_8f2a
00:00.0  plan · 3 steps
00:00.4  tool: search_db · 120ms
00:01.1  guardrail: pass · policy v4
00:01.2  tool: send_email · HELD → approved
00:02.0  done ✓
CONTROL & COMPLIANCE
Policy enforcement
Declarative rules per agent, versioned and tested like code. Changes ship through review.
Human-in-the-loop
Hold any action for approval and route it to the right person — with full context attached.
Audit & replay
An immutable log of every run, retained for compliance and replayable step by step.

Put your agents
on rails.

Get API access Book a demo →
EchoNest-Agent · training format

Rails the 1.26B model is learning

While production still serves language v2, the EchoNest-Agent 1.26B run trains on structured chat + tool traces so the model can stay on rails for agentic workflows.

Chat / reasoning template

### Instruction
{user goal}

### Reasoning
{optional short chain-of-thought}

### Response
{final user-visible answer}

Tool call / result

<tool_call>{"name": "calculator", "arguments": {"expression": "17*19"}}</tool_call>
<tool_result>{"name": "calculator", "ok": true, "output": "323"}</tool_result>

Training mix includes calculator, read_file, list_dir, and write_file stubs over a fake sandbox/ prefix, plus Alpaca instruct and short Python. Source of truth for the protocol: llms/echonest2/AGENT_RAILS.md.