Safety
AI that asks before it acts.
Every LintLabs system runs on the rules we run our own company on. Agents draft; people approve. Each agent gets only what its job needs. And software checks the rules every night — so safety is a mechanism, not a promise.
Six rules
Written down, then enforced by software.
A rule without a mechanism is a wish. So every rule below names the mechanism that enforces it.
01
A person approves what matters
Agents prepare quotations, invoices, sales orders and payments — only a person can submit or send them. Client messages are drafted by agents; any new kind of message, and anything about price, goes to a person first.
Enforced by System permissions: the one role that can submit or send money documents is held by a person, not an agent — and a nightly check fails if an agent ever gains it.
02
Each agent gets only what its job needs
Every agent has its own account, scoped to its job. A drafting agent works through a gateway that will not submit, cancel, email or delete on its behalf.
Enforced by Per-agent accounts, company-locked permissions, and a business-system gateway that refuses submit, cancel, email and delete.
03
Secrets live in a vault
Production passwords and API keys live in a self-hosted, encrypted vault and are delivered to a service when it starts — not pasted into code, documents or chat.
Enforced by One vault as the source of truth, server secrets sealed until run time, and nightly scans that fail if a plaintext secret appears on a server.
04
Our agents reach AI models through one gateway
Our agents call AI models only through a gateway we operate, which caps spend per key and records what was called, when and by which agent.
Enforced by A LiteLLM gateway with per-key budgets and a spend log, and a managed model catalogue that agents cannot change themselves.
05
Watched around the clock
More than 100 monitors watch the systems, and alerts reach our founder. Safe fixes — like restarting a stuck service — happen automatically and are logged. Anything irreversible waits for a human.
Enforced by A “two-way door / one-way door” rule: reversible actions run and are logged; deletions, outages and money need an approve link.
06
The rules are checked every night
Our policies are written as data and compiled into each agent’s configuration, so behaviour follows the written rule. Every night, about 70 automated checks compare reality with the rules.
Enforced by Policy files compiled into agent configuration, an enforcement plugin on every tool call, and nightly audit checks that open a task on any failure.
Tested on ourselves
We run LintLabs this way first.
The same approvals, access rules and nightly checks govern the AI agents that operate LintLabs itself. Clients get a setup that has already been through our own audits.
- ~10 AI agents at work
- Each with its own account, scoped to its job, under one founder.
- 12 policies, enforced in code
- Rules are data, compiled into every agent’s configuration.
- ~70 automated checks every night
- Reality is compared with the rules; any failure opens a task.
- 0 money documents an agent can send
- Agents draft quotations, invoices and payments. Only a person submits them.
Your data
Least data, named owners, clean exits.
We handle personal data under Singapore’s Personal Data Protection Act (PDPA). In practice that means:
- Only what the workflow needs. We connect to the systems and fields a job requires, read-only first.
- You stay the owner. Your records stay in your systems of record; our agents work on them under your approval.
- Credentials in a vault. Production credentials live in an encrypted, self-hosted vault and reach a system only when it runs.
- Model calls are visible. Our agents reach AI models through one gateway we run, so we can see, cap and audit what is sent — and choose the provider per workflow.
- Where things run — plainly. Our main servers are in Singapore, with logs and some client systems in the EU. AI requests go to external model providers, currently DeepSeek and Alibaba Qwen; we tell you which ones your workflow uses.
- Clean exits. When an engagement ends, we remove the access you gave us.
Answers
Safety questions
How does LintLabs keep AI agents safe?
Agents draft; people approve. Only a person can submit or send quotations, invoices and payments, and new kinds of client message or anything about price go to a person first. Each agent has its own scoped account, our agents reach AI models only through a gateway that caps and records spend, and about 70 automated policy checks run every night.
Can an AI agent send emails or invoices on its own?
Not invoices, quotations or payments: the permission to submit or send money documents is held by a person, and a nightly check fails if an agent ever gains it. Agents draft client messages; only a routine kind of message that a person has already approved several times can go out automatically, and anything new or about price always waits for a person.
How do you handle our data?
We handle personal data under Singapore’s Personal Data Protection Act (PDPA). We connect only to the data a workflow needs, your records stay in your own systems, production credentials live in a vault, and each agent is limited to the systems its job needs. Our main servers are in Singapore, with logs and some client systems in the EU, and AI requests go to external model providers — currently DeepSeek and Alibaba Qwen.
Ask us how it would work for you.
Bring the workflow and your toughest question about risk. We’ll show you exactly where the human approvals sit.