Catch AI agent failures that look like success.
Detect duplicate actions, conflicting updates, and false success claims across your business systems. Start in watch-only mode in minutes.
Starts watch-only (never blocks day 1) · Works with LangGraph, Temporal, OpenAI Agents & MCP
you can’t afford to break
- Stripe Refunds
- Shopify Orders
- NetSuite / QuickBooks
- Zendesk / Gorgias
- Salesforce / HubSpot
From agent call
to proven result.
Two agents, one refund. Only one reaches Stripe.
Sample agents make real mistakes in a sample workspace, and you watch Lumni catch them.
In its folder. Lumni only watches: you see what it would stop, and nothing is blocked.
When the record shows it only stops real problems, turn it on per action.
The dangerous parts of agent actions, handled.
Retries, stale reads, people and agents on the same record, tools that say OK when nothing happened. Plain rules catch each one before it costs you.
Stops the repeats
Retries and second agents reach for the same refund. The first goes through; the copy never reaches Stripe.
Makes it re-read
Data changed after the agent looked. It reads again before it writes.
People go first
A rep is in the ticket. The agent waits, then goes on the fresh version.
Approvals in Slack
Risky actions wait for a yes, with what you need to decide.
Read-back, not trust
“Done” is checked in the app itself. Half-finished work becomes an incident.
The real console. Click anything.
A sample company’s day with Lumni, every screen live. Nothing here touches a real account.
It’s just your code.
One check in front of the refund, the email or the CRM write. Plain, deterministic rules answer it. No model decides, so it adds no LLM latency.
1const d = await lumni.checkAction({2 actor_id: 'support-agent', work_id: ticket.id,3 operation: 'stripe.refund',4 effect_key: `refund:${order.id}`,5});6if (shouldExecute(d)) {7 await stripe.refunds.create({ charge });8}
{
"decision": "stop",
"reason": "same effect already ran",
"effect_key": "refund:4821",
"first_seen": "2:14:09 PM",
"decided_by": "rule, no model"
}The same effect twice, data that changed, someone on the record, or a rule you set.
Adapters for LangGraph, OpenAI Agents, Temporal and MCP.
The SDKs, the CLI and the skill are Apache-2.0.
Each line is a decision from a plain rule, made before the action reached a real system.
Twelve ways agents go wrong. One place that handles them.
A retry after a timeout refunds again.
Two agents reply to one customer.
Model calls for work already done.
Acts on data someone just changed.
The customer changed their mind; the agent didn’t.
Ticket closed before the refund landed.
You fix a record; an agent overwrites it.
Many agents edit one account at once.
One agent cancels while another upgrades.
The tool said OK. The app has nothing.
Charged the card; the deal never closed.
One task fans out to forty agents.
See what it would have caught. Then decide.
Watch-only never blocks. It records every action it would have stopped, and after two days you get a risk report with dollar amounts, systems and the exact records.
- Every repeat, stale write and collision, with what it would have cost
- Which actions to protect first
- Something you can forward to your CFO
- Tue 2:14 PMWould stop: repeatStripe refund $840 · Order #4821Retry after a timeout
- Tue 4:02 PMWould hold: person firstGorgias ticket #1183Human rep editing
- Wed 9:41 AMFlagged: half-finishedNetSuite AP invoice #9021Agent said done; not in NetSuite
- Wed 11:20 AMWould re-read: staleShopify order #10342 · cancelOrder changed after read
Priced for agents
that move money.
Start free in watch-only. Pay when Lumni protects live actions, or have us wire it into your stack and prove it on your own traffic in week 1.
Watch-only
See what your agents would break.
- 1,000 events/mo
- Watch-only: records, never blocks
- SDKs (TypeScript, Python, Go) + MCP gateway
- Coding-agent skill
- The full console and action record
Production
Protect live money-moving agents.
- Pay per event, cheaper past 1M
- Stops repeats, stale writes and collisions
- Slack approvals for risky actions
- Read-back verification in the app
- Unlimited people and agents
Enterprise
We wire Lumni into your stack in 48 hours.
- Hands-on integration: LangGraph, Temporal or MCP
- 48-hour Shadow-Mode Risk Report
- Custom read-back verifiers: Stripe, Shopify, ERP, CRM
- A dedicated Slack channel with us
One event is one action an agent asks about before doing it: a refund, an email, a CRM write. Reads, dashboards and people using the console are free.
Running it on your own servers? The source is on GitHub · self-hosting guide
No. New workspaces only watch: Lumni records what it would have stopped and lets everything through until you turn protection on, per action.
Wiring it into a LangGraph or Temporal stack? Book 15 minutes and we’ll look at your riskiest action together.