Claude Partner NetworkOfficial Anthropic partnerGovernance

Autonomy with control, evidence and accountability.

How Luby rolls out agents safely: autonomy levels, human in the loop, guardrails, continuous evaluation, audit and privacy.

Principles

An agent in production is a business responsibility, not an experiment.

Before the first line of code, we define who owns the outcome, how far the agent can act on its own and how we'll measure whether it deserves to keep running. These four principles apply to every agent Luby puts into production.

01

A person owns the outcome

Every agent has a business owner. The agent executes; accountability for the decision stays with the person your company designates.

02

Autonomy proportional to risk

The more irreversible or expensive an action, the closer it stays to a human. Autonomy is expanded with evidence, never by default.

03

Everything is auditable

Every agent decision is logged with the evidence behind it, the tools it used and who approved it.

04

Measure before you scale

A business metric and a kill criterion are set during discovery. If the agent doesn't deliver, it gets switched off.

Autonomy levels

From assistant to autonomous, one step at a time.

We rate every agent on five levels. By default, no Luby agent goes beyond Level 3 on high-impact actions: the agent runs the routine, and exceptions and approval thresholds go to a person.

  1. L1
    Assistant

    Answers and suggests. Doesn't execute anything in your systems.

  2. L2
    Copilot

    Prepares the complete action (draft, entry, reply) and a person confirms every execution.

  3. L3
    Supervised Luby standard

    Runs the routine on its own and escalates exceptions, approval thresholds and low-confidence cases to a person.

  4. L4
    Audited autonomous

    Runs end to end; people audit by sampling. Only after months of evidence at Level 3.

  5. L5
    Fully autonomous

    No human review. We don't recommend it for irreversible, financial or customer-facing actions.

Human-in-the-loop

People step in where the decision matters.

Making someone approve everything wipes out the gain; having no one involved is a risk. We design the exact intervention point in each process.

  • ✓
    Approval thresholdsAmounts, discounts or exceptions above the policy limit go to the right approver, with the recommendation and the evidence ready.
  • ✓
    Confidence-based escalationWhen the agent's confidence drops below the threshold (unreadable document, conflicting data, a new kind of case), it stops and asks instead of guessing.
  • ✓
    Double checks on irreversible actionsPayments, customer communications and critical record changes require explicit confirmation.
  • ✓
    One governance queueA single queue holds everything waiting on a human decision, with deadline, context and a history of who approved what.
  • ✓
    Sample-based reviewEven for automated cases, a sample is reviewed by people and feeds the continuous evaluation.
Regulated decisions

Where the law already has a say, people decide.

Some uses of AI are regulated in the US. In these areas Luby agents prepare the work and document it; a person makes and owns the decision, and your counsel signs off on the rules before go-live.

  • !
    Hiring and promotionScreening and ranking agents support recruiters and never reject candidates on their own. Rules such as EEOC guidance, NYC Local Law 144 bias audits and state AI hiring laws are reviewed with you before use.
  • !
    Credit, leasing and tenant screeningAgents assemble the file and the rationale; approvals, adverse action notices and fair lending or fair housing reviews (ECOA/Reg B, FCRA, Fair Housing Act) stay with your team.
  • !
    Insurance claims and health coverageClaims and prior authorization agents prepare recommendations; coverage denials are made by licensed people under your state's rules.
  • !
    Calls, texts and collectionsOutbound calls and texts follow consent and contact rules (TCPA, FDCPA and Reg F, state laws), with templates approved by people.
  • !
    Protected health informationProjects that touch PHI run under the agreements HIPAA requires, with minimized and masked data.
Guardrails and control

The agent only does what it's authorized to do.

Typed toolsA closed list of actions

The agent works in your systems only through declared tools (via MCP or API) with validated parameters. No free access to screens or databases.

Least privilegeSegregated credentials

Each agent has its own credentials with minimum permissions, stored in a vault and rotated.

LimitsThresholds, volume and cost

Limits by amount, number of actions and run cost. When a limit is hit, the agent stops and alerts.

Kill switchImmediate shutdown

Any agent can be paused instantly by an authorized person, without waiting on the technical team.

Prompt injectionExternal content is data, not orders

Emails, documents and web pages the agent reads are never treated as commands. Sensitive actions require a trusted source.

Output validationChecks before acting

Replies and entries pass business rules and format checks before they reach a customer or a system.

Continuous evaluation (evals)

An agent only moves up a level with evidence.

We evaluate an agent the way you'd evaluate a new hire: with real cases, clear criteria and follow-up after the start date.

  • 01
    A reference set from your processReal, anonymized cases with the expected answer defined by your team. It's the agent's yardstick.
  • 02
    Business and quality metricsAccuracy, time and cost per run, alongside faithfulness to the source, correct tool use and no made-up answers.
  • 03
    Automated grading calibrated by peopleAutomated graders speed up the analysis and are checked periodically against expert judgment.
  • 04
    Regression on every changeEvery change to instructions, tools or model runs against the reference set before it reaches production.
  • 05
    Adversarial testingWe simulate malicious inputs, out-of-pattern data and manipulation attempts before go-live.
  • 06
    Production monitoringDay-to-day samples are evaluated continuously; a quality drop triggers an alert and, if needed, lowers the autonomy level.
Observability and audit

You see what the agent did, why and what it cost.

Trace

Every step logged

Summarized reasoning, tools used, data consulted and the result of every run.

Audit

Immutable decision log

Decisions and approvals stored with evidence, date and owner, ready for internal or external audit.

Cost

Cost per run

Model and infrastructure usage measured per agent, per process and per internal customer.

Alerts

Deviations raise their hand

Falling accuracy, more exceptions, or latency or cost out of range alert the agent's owner.

Security and privacy

Your company's data stays yours.

Luby is certified to ISO/IEC 27001 (information security) and ISO/IEC 27701 (privacy). Every agent inherits that management system.

ISO/IEC 27001ISO/IEC 27701
  • ✓
    Privacy by designPurpose, data minimization and consumer rights under CCPA and state privacy laws (and HIPAA, where PHI is involved) built into the agent's process.
  • ✓
    Masking of personal dataSensitive data is masked before it reaches the model whenever the task allows.
  • ✓
    Models under contractWe agree with you on model providers with zero-retention and no-training terms for your data, or models running in your own environment.
  • ✓
    In your cloud, if you preferDeployed in your own cloud account or an isolated environment, encrypted in transit and at rest.
  • ✓
    Isolation per clientKnowledge bases, memory and credentials are never shared between clients.
Responsible rollout

From pilot to operations, without a leap in the dark.

01

Discovery

The real process, rules, risks, business metric and kill criterion defined with the process owner.

02

Shadow mode

The agent runs alongside the team without acting. We compare its decisions with your people's.

03

Copilot, then supervised

The agent starts executing with human approval and, with evidence, takes over the routine and escalates exceptions.

04

Operate and expand

Continuous monitoring, periodic autonomy reviews and expansion to the next processes.

From the catalog

Agents that govern the other agents.

Part of the governance is also done by agents, always under human supervision. The technical foundation behind these controls is on Technology.

Cross-functional · Governance

Multi-Agent Orchestrator

Coordinates specialist agents, delegates tasks and consolidates results under supervision.

All industriesAutonomy 3 · Supervised
Learn more →
Cross-functional · Governance

Memory & Context Agent

A long-term memory layer: captures interactions and personalizes the other agents over time.

All industriesAutonomy 1 · Assistant
Learn more →
Legal & Compliance · Privacy & Compliance

Privacy & Data Rights Agent (CCPA)

Keeps your personal data inventory current, answers consumer data requests within legal deadlines (CCPA and state privacy laws) and flags processing without a stated purpose.

All industriesAutonomy 2 · Copilot
Learn more →
References

The frameworks that guide our governance.

The practices on this page follow public standards and references for security, privacy and AI risk management.

Companies that trust Luby

  • Toyota
  • PwC
  • Siemens
  • Bridgestone
  • LexisNexis
  • Sunwest Bank
  • Papaya
  • Foxtrot
  • Usend
  • Inter&Co
  • Abbiamo
Testimonials

The people who work with Luby speak for us.

“The way they understood our vision and turned it into a working product is truly impressive.”
Rafael Rjeille CordeiroCEO, Yangflow
“Everyone is professional, agile and flexible. They are open to feedback and keep improving their process.”
Jason MeltzerCTO & Co-Founder, Papaya
“Luby is an innovative company that is committed to its projects. I feel confident taking on the most challenging work with them as partners.”
Wilson Baraban FilhoCTO, Inter&Co
Why Luby

23 years of shipping software. A journey designed to remove risk.

Luby has taken more than 1,350 projects to production for companies in the US and Latin America. North American leadership in Miami, senior nearshore engineers whose working day overlaps yours. About Luby →

23+years in production software
1,350+projects delivered
220+engineers and product specialists
4.8/537 verified reviews on Clutch
ISO 27001 · 27701certified security and privacy
Before the build

Discovery with a written exit

Scope, success metric and kill criterion agreed in writing before any code. If the case doesn't hold up, you stop there.

Before it acts

Shadow mode first

The agent runs next to your team without touching anything. You compare its decisions with your people's before it goes live.

While it runs

Your people approve

Supervised by default: every high-impact action has an approver, an audit trail and a kill switch your business team controls.

Your environment

Your cloud, your data

Deployed in your cloud or an isolated environment, models under no-training terms. AWS and Google Cloud certified partner.

Talk to an AI architect

Talk to an architect about rolling out agents safely

Bring your security, compliance or IT questions. We'll walk through how a first agent would run in your environment.

  • 1
    A senior architect, not an SDR30 minutes with someone who has shipped agents to production.
  • 2
    Your top 3 agent opportunitiesRanked by payback, with the systems and approvals each one needs.
  • 3
    No commitmentYou keep the shortlist whether or not you work with Luby.
Book a time directlyRun the free website assessment