retrace

// enforcement

Block bad actions before they run

A pre-call decision engine that decides allow, block, or hold for an agent action before it executes, so runaway loops and budget blow-outs are stopped at the source, not diagnosed after the fact.

Start freeRead the docs ā†’

Pre-call gate

Before a model or tool call executes, a single check returns allow, block, or hold, the action only proceeds if the policy permits it.

Fails safe

Enforcement is treated like billing and auth: if a decision can't be made, the action is blocked rather than allowed through.

Ceilings, loop breakers, debounce

Cap spend per run, break repeated identical actions, and debounce rapid duplicate calls, individually or combined into one policy.

Hold for approval

High-risk actions can be parked in an approval queue instead of blocked outright, with versioned policies so every change is auditable.

Drop-in gateway

Route model calls through a provider-compatible gateway to enforce policy without changing your client, and your provider key is never stored or logged.

Offline SDK ceilings

The SDK enforces local per-run step, token, and spend ceilings with zero network calls, raising a typed RetraceEnforcementError instead of silently skipping work.

agent.py
from retrace import RetraceEnforcementError

try:
    run_agent(prompt)
except RetraceEnforcementError as e:
    # e.verdict is "block" or "hold"
    print("stopped by enforcement:", e.verdict)

A blocked or held action stops the call with a typed error.

check.http
POST /api/v1/enforcement/check
{
  "run_id": "run_123",
  "tool_name": "send_email",
  "proposed_usd": 0.20
}

ā†’ { "verdict": "allow" | "block" | "hold" }

Or ask the gate directly before running any action.

REST API

  • POST /api/v1/enforcement/check: allow / block / hold decision

More of the platform

// recordRecordOne decorator turns every run into a regression test// fork & replayFork & ReplayFork the failed step and re-run, debug the cause, not the symptom// prove the fixProve the FixRe-run a fix and get a verdict before you ship// eval gatesEval GatesFail the build when agent behavior regresses// guardrailsGuardrailsHalt a run when it blows a budget, loops, or stalls// failure taxonomyFailure TaxonomyAuto-classify failures so you fix patterns, not one-offs// sessionsSessionsTrace multi-agent hand-offs so failures aren't a black box