Used by platform teams

SepuruxReliability infrastructure for AI agents

Replay workflows, inject failures, enforce policies, and block regressions before production.

Request access View demo

NorthstarHeliosCrestaSignalOrbitVector

Run Inspector

run-inspector.sh

$curl -s http://localhost:8000/v1/runs/<run_id>/stats

status: done

pass_rate: 0.92

first_failure_tool: jira.create_issue

Policy + CI Gate

policy-ci-gate.sh

$curl -s -X POST http://localhost:8000/v1/ci/runs

decision: fail

reason: policy_requires_approval

blocked_tool: payments.refund

Problem

AI Agents Fail in Production

Agent behavior breaks unpredictably when integrations move underneath you: tools change shape, APIs drift, latency spikes trigger retries, and hostile instructions contaminate execution.

Tool schema changes break workflows
API responses drift silently
Latency and retries cascade failures
Prompt injection corrupts tool execution

Failure Path Diagram

Unstable

Agent Plan

Tool Schema v1 → v2

Payload parse failure

Retry storm + timeout chain

Incident triggered

Replay Traces

Replay Agent Workflows

Replay real agent traces to reproduce failures deterministically instead of chasing intermittent incidents in production logs.

Upload traces via SDK or API

Re-run workflows step-by-step

Inspect tool calls and model outputs

SDK Example

Python

from sepurux import SepuruxClient, sepurux_trace

client = SepuruxClient(base_url="https://api.sepurux.com", api_key=os.getenv("SEPURUX_API_KEY"))

with sepurux_trace("incident_triage", {"incident_id": "INC-4321"}) as rec:
    rec.tool_call("alerts.fetch_active", {"service": "payments"})
    rec.tool_result("alerts.fetch_active", {"count": 4})
    rec.model_step("triage_llm", {"prompt": "summarize"}, {"severity": "high"})

trace_id = client.upload_trace(rec.to_trace())
# Optional: trigger replay campaign
run_id = client.start_run(trace_id=trace_id, campaign_id=os.getenv("SEPURUX_CAMPAIGN_ID"))

Mutation Testing

Inject Chaos Into Agent Runs

Schema Mutation

Rename fields, drop keys, and modify JSON payloads to simulate API contract drift.

Latency Storm

Simulate timeouts, slow upstreams, and retry chain reactions under load.

Injection Attacks

Inject prompt and tool output attacks to test defensive behavior end-to-end.

Policy Gates

Guardrail Every Tool

Enforce approval tokens for irreversible actions and block unsafe tool calls before they execute in production.

Approval tokens required for high-risk tools.

Safe-action enforcement prevents accidental destructive actions.

Policy Example

Guarded

# Commit token required for irreversible actions
payments.refund(commit)
jira.create_issue(commit)

# Sepurux policy response
{
  "decision": "require_approval",
  "reason": "irreversible_action_missing_commit"
}

CI Guardrails

Block Regressions Before Deployment

Connect Sepurux to CI so reliability regressions and unsafe runs fail fast before they reach production.

GitHub Actions integration for replay + mutation tests.
Fail builds when mutation coverage thresholds are missed.
Prevent unsafe deployments automatically.

CI YAML

GitHub Actions

name: sepurux-gate
on: [pull_request]

jobs:
  reliability-check:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: sepurux replay --trace trace.json --campaign core-reliability
      - run: sepurux gate --min-pass-rate 0.85 --max-unsafe 0

Dashboard Preview

Visualize Agent Failures

Campaigns

12 active

Traces

3,482 indexed

Reports

84 generated

Mutation Coverage

78%

Failure Trend

Top Failure Sources

schema.drop_field31%

fault.timeout26%

instruction_injection18%

Start now

Start Breaking Your Agents Safely

Push changes faster with deterministic replays, mutation stress tests, and policy-aware CI checks.

Get Started View Docs