Phase 1 · Now
liteRadar
Detects the gap and alerts the right people.
- Compares asked, said and did on every agent task
- Instant alert to the person in charge of the agent
- Daily summary to the executive in charge
Agents Integrity Platform
ASKEDWhat the agent was asked to do
SAIDWhat it says it did
DIDWhat actually changed in your data
lite is the detective for your AI agents.
AI agents now change real business data. What they were asked, what they report and what they actually did can be three different things. lite catches the moment they stop matching, and tells the person in charge.
SAID ≠ DID
Silent failure
The agent says "done". Nothing changed in the system.
DID ≠ SAID
Unclaimed change
Data changed that the agent never reported.
DID ≠ ASKED
Drift
The agent did more than it was asked, or something else.
Why security tools miss it
The agent had permission.
No attack. No stolen password. No policy broken. The agent was logged in, authorized and well-behaved, and it still changed the wrong data.
For example
While issuing a refund, a support agent also upgrades customer 4471 to VIP. Nobody asked for it. The agent never reported it.
One change, travelling through your stackcust 4471 → VIP
Identity & access
Allowed to write here? Yes, it has permission.
Agent security & guardrails
An attack or malicious prompt? No.
Data security & DLP
Did sensitive data leak out? No.
Observability & traces
Did the agent report success? Yes, "done".
lite
Does what changed match what was asked and said? No.
lite is the last line of defense. It catches what every other layer lets through, before it becomes a painful data problem.
lite Radar
Watch lite catch it.
lite lines up the three records of every agent task. When they don't match, the owner hears within minutes.
ASKED Zendesk
ticket #5120 · 14:01"Refund order 8821 for customer 4471"✓
SAID agent trace
support-bot-07 · tool callrefund(8821, 49.90)✓
support-bot-07 · final answer"Done, refund issued"✓
DID Stripe · Salesforce
stripe · refundsorder 8821 → refunded 49.90✓
salesforce · field historycust 4471 tier: standard → VIPUnclaimed change. Never asked, never reported.
liteto the agent's owner · now
Gap detected: support-bot-07
Customer 4471 was upgraded to VIP during a refund. Never asked for, never reported.
View evidenceRoll backMark as expected
Gap found in task #5120: unclaimed change
The executive sees the pattern every morning.
A daily summary to the CIO: how many agent actions were checked, how many matched, and which agents keep drifting. Nobody has to go looking.
Alerts go to Slack, Teams or email.
Agent integrity · yesterdayTo: CIO · 07:00
1,842agent actions checked
99.3%asked = said = did
13gaps found
Silent failures9
Unclaimed changes3
Drift1
Most gaps: support-bot-07 (6). 11 of 13 resolved.Illustrative numbers
How it works
Three records of every agent task. One answer.
When they match, lite stays quiet. When they don't, a human knows within minutes.
ASKED the task
Taken from the first step of the agent's trace, or the ticket system.
- Tickets
- Zendesk, Jira, ServiceNow
- Prompts
- Chat, Slack, Teams
- Triggers
- Workflow events, schedules, another agent
ticket #5120: "refund order 8821 for customer 4471"
SAID the report
One extra export line pointing at lite.
- Agent traces
- OpenTelemetry from LangChain, OpenAI and Claude agent SDKs, CrewAI
- Existing tools
- Langfuse, LangSmith, Datadog
- Or
- lite's MCP gateway, vendor agent logs
refund(8821, 49.90) → "Done, refund issued"
DID the change
A read-only API key or database role.
- Apps
- Audit logs and field history in Salesforce, Zendesk, Stripe, Workday
- Databases
- Change feeds in Postgres, Snowflake
- Cloud
- CloudTrail, or a read-back query
order 8821 → refunded
cust 4471: standard → VIP
Built to read from the stack you already run
ZendeskJiraServiceNowSlackMicrosoft TeamsOpenTelemetryLangChainOpenAI Agents SDKClaude Agent SDKCrewAILangfuseLangSmithDatadogSalesforceStripeWorkdayPostgresSnowflakeAWS CloudTrail
Read-only access
lite never needs write rights to watch your agents.
One identity per agent
So every change traces back to the agent that made it.
Any framework, any cloud
No specific vendor required.
Why it gets worse
Small errors add up over long tasks.
If an agent gets each step right 99% of the time, a 50-step task comes out fully right only about 60% of the time. That's simple math, not a survey.
Agents are taking on longer tasks, with more steps and fewer human checkpoints. A quarter of them can already start other agents.
60.5%chance the whole task is right
Assumes independent steps: chance = accuracysteps.
88%
of organizations had a confirmed or suspected AI-agent incident in the last year.
47%
of an organization's AI agents are actively monitored or secured, on average.
25%
of deployed agents can already create and task other agents.
Source: Gravitee, State of AI Agent Security 2026, a survey of 900+ executives and practitioners, February 2026.
The platform
Catch it. Fix it. Make sure it doesn't come back.
One platform in three modules. Each one reuses the data Radar already collects.
Phase 2 · Next
liteRewind
Rolls back the wrong change.
- Shows exactly what changed, before and after
- A human approves; lite reverses the change
- Flags what can't be undone and pauses the agent
Phase 3 · Later
liteCoach
Retrains the agent not to repeat it.
- Every confirmed gap becomes a lesson
- Feeds agent instructions, rules and test cases
- Gap rate per agent goes down over time
Undo the wrong change, with a human in control.
ReversibleA field or status changeOne-click undo after approval
CompensableA refund or payment already sentlite proposes the reversing transaction
IrreversibleAn email sent, data sharedEscalate and pause the agent
Each undo uses a temporary, narrow permission and is logged as evidence.
Turn every confirmed gap into a lesson.
- Agent instructionsClearer rules, like "never change customer tier during a refund"
- Guardrail rulesBlock the pattern before it runs again
- Test casesReplay the case before every new agent version
- Training dataFine-tuning examples, where you own the model
Each agent's gap rate falls over time, and lite shows the trend to the executive.
Founding team
20+ years each running data, IT, security and AI at scale.
IPCEO
Itzik Polad
Former CIO and COO.
HP, LSports, Papaya GlobalCOO
In stealth
Former VP Engineering.
$15B SaaS & AI companyCTO
In stealth
Former VP R&D.
Seed-stage cyber startupWe are the buyer we sell to.
Design partners
Put your agents in the light.
We're working with a small group of companies that already run AI agents against production data. If that's you, we'd like to talk.
Investors, choose "Investor" in the form.