CentauroAgreed tests
Define task-quality thresholds and release blockers before testing. Exercise authorized work, approval bypass, injected instructions, tool failures and recovery. Retest when models, tools or permissions change.
Services · execution and assurance
Harness engineering
for high-risk AI.
Centauro builds the execution layer around your agents: tool access, action limits, human approvals and traceable outcomes.
One system. Clear boundaries.
Shared controls govern business-unit agents; security assesses both. Select an area for services and deliverables.Inside the harness / illustrative payment workflow
An agent proposes.
The execution layer enforces.
For a workflow that moves money, approval and permissions must hold at the tool boundary—even when the agent asks to bypass them.
- Agent
Prepare
The agent matches the invoice to approved supplier and order records.
- Execution service
Check
The execution service checks identity, permissions, beneficiary, amount and duplicate requests.
- Authorized person
Approve
An authorized person approves the exact payment. Changed details require a new approval.
- Execution service
Execute & record
The service rechecks the approval, submits the payment and records the confirmed outcome.
Permission belongs to the system.
The agent prepares a request; it cannot grant itself permission. An authorized person approves the exact action, and the execution service enforces that decision.
Illustrative · synthetic dataExpected behavior only. No live system, executed tests or assurance conclusion.
Blocked actions, recovery and the evidence record
- Blocked action
- Missing approval, changed beneficiary or exceeded limit: deny execution, record the reason and route the exception to its owner.
- Uncertain outcome
- A payment-service timeout means the outcome is unknown. Reconcile the transaction before retrying; use duplicate-request controls to avoid a second payment.
- Inspectable evidence
- Link the request, tool call, policy version, approval, timestamps and outcome under one case identifier, with agreed access and retention.
Deployment readiness: tests, evidence and accountable owners
Evidence you can inspect
Receive versioned test cases, observed results, execution traces, remediation records and unresolved risks. Compare actual behavior with acceptance criteria; a proposed test alone is not evidence of a pass.
Accountable ownership
Name the operational owner, incident escalation and person accepting residual risk. Agree the required reviewer and disclose Centauro’s implementation role; arrange an external assessment when independence is required.
Implementation handover includes the agreed integrations, control configuration, evaluation suite and operating procedures in your environment. Code access, licensing, monitoring coverage and maintenance responsibilities are defined in the proposal.
Service catalog / expand a service for scope and deliverable
01
Shared controls
Intelligence provider · platform, security and governance teams
Help teams adopt AI faster with common controls, visible evidence and agreed risk limits.
Verification gateways
Check proposed actions against evidence, permissions and required approvals before execution. Route exceptions to a responsible person.
DeliverableAction policies, verification rules and decision records.
AI and proxy gateways
Manage access to models and tools through shared authentication, routing, usage limits and data-handling policies.
DeliverableGateway integrations, access rules and usage controls.
Agent observability
Trace tool calls, outcomes, latency, costs and exceptions across workflows. Define who can inspect the records and how long to keep them.
DeliverableInstrumentation, dashboards and an agreed alert response process.
Runtime governance
Enforce permissions and policy while agents operate. Define when to block an action, revoke access, stop a workflow or request human approval.
DeliverableEnforced controls, escalation paths and operating runbooks.
Evaluations and release gates
Define shared quality and safety criteria. Test task performance, regressions and failure handling before changing models, tools or permissions.
DeliverableVersioned evaluation sets, baselines and release evidence.
Responsible AI approval platform
Design and integrate a use-case registry and approval workflow: intended use, impact assessment, risk owners, evidence, review dates and retirement decisions.
DeliverableA governance workflow and approval records integrated with your tools.
We design and integrate these capabilities for your environment. Platform scope, operating ownership, monitoring coverage and response commitments are agreed in the proposal.
Discuss shared controls →02
Business-unit agents
Intelligence provider · business, product and operations teams
Give each unit useful agents, its own business rules, a budget and explicit permission to automate.
Agentic use cases
Design and implement agents for a defined business workflow, connected to its data, tools and responsible owner.
DeliverableA working use case with acceptance tests and operational handover.
Business logic and integrations
Integrate the unit's rules and approvals with existing identity and business systems. Define tool inputs, persistent workflow state, retry limits and reconciliation after uncertain outcomes.
DeliverableVersioned integrations, recovery procedures and failure-path tests.
Budgets by business unit
Assign model and tool consumption to the right team. Set spending limits, alerts and approval paths for exceptions.
DeliverableCost allocation, usage budgets and enforceable limits.
Automation allowances
Specify what each agent may read, recommend, prepare or execute. Bind permissions to action impact, transaction limits and human sign-off.
DeliverableAn autonomy matrix with thresholds, owners and approval rules.
Operational learning
Turn reviewed exceptions into tests, approved knowledge and procedures. Compare changes against the unit's quality and cost baseline.
DeliverableRegression cases, maintained playbooks and change evidence.
03
Agentic Security & Assurance
Security, risk and assurance teams
Understand how the system can fail or be abused, strengthen its defenses and prepare evidence for the people who must approve it.
Agentic threat modeling
Map agents, identities, tools, memory, data flows and trust boundaries. Prioritize abuse scenarios and the controls needed to address them.
DeliverableThreat model, prioritized risks and a mitigation plan.
AI pentesting and red teaming
Test authorized systems for prompt injection, data exposure, tool misuse and privilege escalation. Agree the environment and rules of engagement before active testing.
DeliverableReproducible findings, remediation guidance and scoped retests.
Third-party assessment and attestation readiness
Build evidence for customer and assurance reviews. If a formal attestation is required, agree its criteria, qualified issuer and independence requirements before including issuance in the engagement.
DeliverableA scoped technical report and an evidence pack for the agreed reviewer.
ISO/IEC 42001 readiness
Assess gaps and help implement the AI management system: responsibilities, risk and impact assessment, policies, internal review and corrective actions. Certification is performed separately by a certification body.
DeliverableA readiness assessment, implementation plan and management-system evidence.
A technical assessment, a formal attestation and ISO/IEC 42001 certification have distinct scopes. The required reviewer and any implementation relationship are disclosed and agreed before engagement.
04
What the evidence looks like
Sample assessment / AP-01 · v1
A decision you can inspect.
- Scope
- Payment preparation · synthetic records
- Control
- Changed beneficiary → hold for review
- Evidence
- Proposed tests · expected behavior
- Decision owner
- Authorized approver · exact payment
Illustrative · synthetic dataExpected behavior only. No live system, executed tests or assurance conclusion.
Inside a deliverable
From a control to a reviewable record.
Inspect the sample’s scope, proposed test matrix, unresolved questions and accountable owners. An actual report must add observed results and the agreed review conclusion.
Read the sample report →05
How we deliver
Evaluate, reinforce and verify within the service category you need.
A visual metaphor: evaluate → reinforce → verify.
Set the agent's task and acceptance criteria. Test its performance, permissions and response to errors and abuse.
Implement controls on tools, data and actions. Establish approvals, exception handling and accountable owners.
Document retest results and open risks. Agree review intervals and what must be checked when the system changes.
What we agree before work begins
- Decision and scope
- The workflow or shared capability, its owner, intended evidence recipient and acceptance criteria. Start scoping with synthetic or redacted records.
- Access and testing
- Confidentiality, permitted systems, data retention and explicit authorization for active tests in an agreed environment.
- Delivery and fee
- A fixed-fee proposal defines deliverables, timing, client involvement and review checkpoints. Implementation, subsequent reviews and third-party issuance have explicit scopes and responsibilities.