Skip to content

AInda Controle

The software layer that makes agentic AI auditable and scalable.

AInda Controle (our control layer) runs on any AI harness, consuming its telemetry and normalizing it into reports intended for corporate auditing. It is not a software license sold in isolation: it is the instrumentation of AInda's service. It is what allows the client to say yes to increasing autonomy for agents at each revalidation, because an agent moves up an autonomy level only when its reliability is measured and proven: reliability before autonomy. The model proposes; a separate control decides before each action.

↑ Authority

  1. Initial authority
  2. Revalidation:Raise
  3. Revalidation:Keep
  4. Revalidation:Raise
  5. Or reduce

Quarters of operation →

Each quarterly revalidation can raise, keep, or reduce the authority. It is not automatic, and the drawing has no scale: no step carries a value.

How it works

Every action goes through AInda Controle

AInda Agentes (our reference harness) and the agents your company already uses follow the same rules: authority is decided before the action, the right person approves on their own channel, and everything is recorded.

Approval request, on each person's channel

People

and their channels

  • Web and mobile
  • Teams
  • Slack
  • WhatsApp
  • Email

Who asks, who approves, who audits

AInda Agentes

the company's team of agents

  • Agents by department
  • Tiered memory
  • Approved skills
  • Isolated execution

Third-party agents

what the company already uses

  • Copilot and Copilot Studio
  • ChatGPT and Codex
  • Claude and Claude Code
  • Gemini and Agentforce

AInda Controle

every action goes through here

  • Authority decided beforehand
  • Human approval
  • Tamper-evident evidence
  • Cost per unit of work
  • Inventory and owner of every agent
  • Immediate containment

Systems

of the company

  • ERP and CRM
  • Documents
  • Internal APIs
  • Data and spreadsheets

AI models

chosen per agent

  • In the client's cloud
  • From the vendors
  • Open-weight, in Brazil
  • On standby

Where it runs

  • AWS, Azure, or Google
  • National cloud or hosting provider
  • The client's data center
  • Dedicated, in Brazil

Data and model residency in Brazil available, at the client's choice

AInda's agents and the ones the company already uses go through the same control before they touch systems and models. When an action needs approval, the request goes back to the right person, on their own channel.

The path of an action

The agent asks

Pay supplier: R$ 50,000

Accounts payable agent

AInda Controle

Checks the authority matrix before the action

Agent limit: up to R$ 5,000

  1. Within the limit

    Executes

    And records

  2. Above the limit

    Human gate

    The approver decides on their own channel; the “yes” and the “no” are recorded

  3. Outside the matrix

    Blocks

    And alerts the agent owner; the block is recorded

Tamper-evident record

Who asked · rule applied · who approved · what happened · what it cost, all tied to the unit-of-work ID

The authority matrix is checked before the action. Within the limit, the agent executes; above it, a person decides; outside the matrix, the action is blocked. All three paths are recorded, including the no.

Components

The six components

  1. 01AI Agent inventory and identity
    AgentOwnerAuthorityReview
  2. 02Behavior observability
    Human intervention
    Out-of-scope actions blocked
  3. 03Agentic FinOps
    BudgetDrift alert
  4. 04Evidence and audit
    Package per process
  5. 05AI Model continuity
    Primary
    Secondary
  6. 06AI agent lifecycle
    ReviewEnd of lifeDeactivation
Schematic illustration, not a screenshot.

  • 01AI Agent inventory and identity

    What it delivers

    A living record of every agent: named human owner, declared authority, scope, data accessed, execution cost, review date, and retirement date. Each agent is a non-human identity, with a named owner and least privilege. It covers shadow AI too: Controle finds agents created in unauthorized personal tools, such as OpenClaw and Hermes, and AInda migrates each one, with its skills, to the governed harness, with owner, authority, evidence, and sandbox. From shadow AI to governed AI.

    Why it matters

    82%of organizations found unknown AI agents running in their own environmentCloud Security Alliance

  • 02Behavior observability

    What it delivers

    Audit and decision trails, human intervention rates, out-of-scope actions blocked, and behavior drifts — not just simple error logs.

    Why it matters

    Only 11%automatically block agent actions that fall outside the authorized scopeCloud Security Alliance

  • 03Agentic FinOps

    What it delivers

    Costs per task, per agent, per process, and per unit of work; budgets with automatic cut-off; cost drift alerts; explicit currency exposure.

    Why it matters

    26% / 36%have real-time visibility of AI cost and direct control over token consumptionKPMG, Global AI Pulse Q2 2026

  • 04Evidence and audit

    What it delivers

    Evidence packs per process, in a format internal, external, and regulatory auditors can use, linking each decision back to its data source.

    Why it matters

    Only 10%maintain both audit trails for AI agents and links from the decisions back to the data sourcesModern Data Survey 2026
    • LGPD, art. 20
    • Res. CMN 4.893/2021
    • Res. BCB 85/2021
    • Res. CFM 2.454/2026

  • 05AI Model continuity

    What it delivers

    Primary models, secondary models with possible sovereign or local hosting, with tested switchover plans.

    AI agent

    Primary model

    In use.

    Secondary model

    On standby. May use sovereign or on-premises hosting.

    Switchover plan

    Tested before it is needed.

    Switch over when

    • The provider suspends access
    • The provider deprecates the version
    • A better alternative appears
    The secondary and the switchover plan exist before they are needed, and the plan is tested — not improvised on the day access is suspended.

    Why it matters

    June 2026access to two frontier models was suspended by a US export directive on June 12 and redistributed on July 1 — model continuity stopped being a theoretical scenarioAnthropic

  • 06AI agent lifecycle

    What it delivers

    Mandatory periodic review, retirement criteria, and a decommissioning process.

    Before production

    1. 01UnderstandAn end that is rarely covered
    2. 02Specify
    3. 03Plan

    In production

    1. 04Production
    2. 05Steady state
    3. 06Revalidate

    Revalidate · every quarter

    End of life

    1. 07DeactivateAn end that is rarely covered
    The seven gates with client participation, in order. Revalidation comes back every quarter, and deactivation is a stage of the cycle, not a cleanup.

    Why it matters

    Only 21%have a formal process for decommissioning agentsCloud Security Alliance

Functional architecture

Where every action goes

  1. Channels

    People and systems

    • Web and desktop
    • Teams and Slack
    • WhatsApp and email
    • APIs and events

  2. AInda Agentes

    The team of agents, optional

    • Agents by department
    • Durable orchestration
    • Isolated execution
    • Memory and knowledge
    • Approved skills

  3. AInda Controle

    Every action goes through here

    • Identity and inventory
    • Authority decision
    • Human approval
    • Evidence ledger
    • Cost and SLA

  4. Connections

    Under Controle's decision

    • AI models
    • Tools and MCP
    • Company systems
    • Third-party agents

  5. Where it runs

    At the client's choice

    • The client's cloud
    • National cloud or hosting provider
    • The client's data center
    • Dedicated, in Brazil

Operations

Spans every layer

  • Lifecycle gates
  • Continuous evaluation
  • Incidents and containment
  • Quarterly review
  • Dashboards by audience
  • Fleet updates
Channels bring people and systems to the agents; every action goes through AInda Controle before it reaches models, tools, and systems; and operations span every layer.

The cycle of an action

  1. 01Request

    From a person, a system, or another agent

  2. 02Identity

    Who acts, on whose behalf, with which owner

  3. 03Decision

    The matrix answers first: execute, ask for approval, or block

  4. 04Execution

    In an isolated environment, with an authorized budget and tools

  5. 05Validation

    The output is checked before moving on

  6. 06Evidence

    Everything tied to the unit-of-work identifier

  7. 07Measurement

    Cost, accuracy, and intervention feed the SLA

↺ Evidence feeds the authority review, the model choice, and skill improvement

Architecture

A three-layer architecture

Our intellectual property does not live in the harness. It lives in portable artifacts. The boundaries between layers are explicit on purpose — that is what makes the portability commitment enforceable.

03

AInda Controle

  • Governance
  • Observability
  • FinOps
  • Evidence

02Execution harnesses

⇄ Swappable, like the model and the cloud

  • AInda Agentes

  • Another harness

  • Client's own infrastructure

01

Portable artifacts

  • Agent Skills
  • MCP servers
  • Evaluation suites
  • AWS Agent Plugins

The company's asset: it survives a change of harness, of model, and of cloud.

  1. 01

    Portable artifacts

    Declarative skills in the open Agent Skills standard, MCP servers and configurations for enterprise systems, per-process evaluation suites, and packaging in the AWS Agent Plugins standard. This layer is the company's asset: it survives a change of harness, of model, and of cloud — and it is what makes each new process cost less than the one before it. Every skill brought in from outside or created in a conversation goes through a governance gate that checks security and compliance and decides to adapt, accept, or reject it, with the reason in the evidence. For an accepted skill, every call is recorded at the tool boundary, through a governed gateway, not by self-reporting.

  2. 02

    Execution harness

    Selected according to the use cases and requirements of the customer or regulator. Depending on the requirements of formal certification, data residency, data processing rules, or corporate identity controls, the answer may require different harnesses.

  3. 03

    AInda Controle

    Governance, observability, FinOps, evidence, and audit trails, designed to run on top of any layer-2 harness.

What you take if you leave

The client takes, in an open format

  • Agents
  • Skills
  • Integrations
  • Inventory
  • Records
  • Authority matrices
  • Evaluations
  • Reports

They can operate with another partner or on their own; the transfer is in the contract from day one

The client loses

  • The partner
  • The AInda Controle dashboard

Not the process, nor the evidence

The open standards we support

  • Agents and tools

    • MCP
    • A2A
    • AG-UI

  • Identity

    • OAuth 2.1
    • OIDC
    • SCIM
    • SAML
    • SPIFFE

  • Authorization

    • AuthZEN
    • Policy as code

  • Telemetry

    • OpenTelemetry
    • GenAI conventions

  • Security

    • OCSF
    • SBOM
    • Signed images

  • Cost

    • FOCUS

  • Events and integration

    • CloudEvents
    • REST
    • Webhooks

  • Evidence

    • JSON Lines
    • Parquet
    • RFC 3161 timestamp

Reference harness

AInda Agentes, your company's team of agents

Every project has AInda Controle. AInda Agentes is optional: it comes in when it is the right choice for the case, alongside our multi-harness policy.

Agents by department, each with an owner, limits, and cost, running on a server in your environment. Company, department, and agent memory; skills approved by the AI Committee; human approval through each person's channel — web, Teams, Slack, WhatsApp, or email. Two regimes: official processes, with a behavior SLA, and supervised free use, with the department's budget and skills. Software engineering stays with the market-leading tools, which AInda Controle governs.

The client's hub is a desktop app: manage agents, respond to approval requests, and follow cost, reports, and alerts in one place. Recurring agents run in the back office of your environment; interactive agents run on the machine of the person using them, with audit trail, sandbox, and authority on top. Approvals arrive as cards, with an effect preview, a risk summary, and a diff, drawn from a closed catalog that the app itself renders: the model only fills in the data, and every button that changes something goes through the authority check and stays in the evidence.

We offer a customizable reference harness for each client. The solutions we develop for you — including skills, integrations, behavior tests, and audit trail — are provided in an open format. This lets you switch harnesses, models, or clouds, leverage corporate agreement benefits, and take them wherever you choose. It's included in our contract.

Data

Where your data lives

By default, AInda Controle is installed in your environment. Your data and evidence never leave your infrastructure; we access it with agent credentials that you delegate and can revoke. If you prefer, we run an instance dedicated to your company, hosted in Brazil. And your data does not train models unless you explicitly request it. Sovereignty in three dimensions, always your choice: data (where the data lives), operational (who operates and accesses it, with your keys), and technological (the essentials under your control, without external dependency).

AInda Controle

Governance, evidence, and authority always belong to AInda, wherever the agents run

  • In your cloud

    Native mode

    AWS, Azure, or Google: the cloud's agent services execute, AInda Controle governs

  • On any Kubernetes

    Portable mode

    The AInda platform, the same on any cloud or data center

  • National cloud or hosting provider

    Autonomous mode

    All it takes is virtual machines, a private network, and storage; we bring the rest

  • Dedicated, in Brazil

    Operated by AInda

    An instance just for your company, in a cloud in Brazil

  • Data and model residency in Brazil: the client's choice
  • Shared environment with client data: never
Execution can stay in your cloud, on any Kubernetes, in a national cloud, or in a dedicated instance in Brazil. In every mode, governance, evidence, and authority stay with AInda Controle.
  • In your environment, by default
  • Access through agent credentials that you delegate and can revoke
  • Or an instance dedicated to your company, hosted in Brazil
  • Your data does not train models unless you explicitly request it

How to choose the mode

  1. 01Do you want your cloud's agent services to do the work?

    Yes

    Native

    AWS, Azure, or Google Cloud; Controle governs on top

  2. No

    02Do you have Kubernetes, in the cloud or in the data center?

    Yes

    Portable

    The full AInda platform, with whatever managed services exist

  3. No

    03Are you on a national cloud, a hosting provider, or your own infrastructure?

    Yes

    Autonomous

    All it takes is virtual machines, a private network, and object storage

  4. No

    04Would you rather not operate the infrastructure?

    Yes

    Dedicated in Brazil

    An instance just for you, operated by AInda; never shared

In every mode: governance, evidence, and authority under the same rules; residency in Brazil at the client's choice.

Harness engineering is the reason we are able to take responsibility for outcomes.

Book an intro callOur method