CentauroOur thesis · growth through learning
Build the capacity
to keep improving.
Experience creates lasting value when an organization can turn what happens into better decisions, tested controls and knowledge it retains.
Growth begins with learning.
Each engagement should leave your team better equipped to investigate a problem, test a response and retain what it learns. Useful experience becomes a capability you can use again.
Unexpected results expose assumptions and weak controls. Improvement requires investigation, a tested response and a record of what changed.
Human judgment. Agentic iteration.
The centaur model pairs people who define the objective, acceptable risk and release conditions with agents that execute tasks and explore alternatives within those boundaries. More iterations create opportunities to learn; the team still needs a way to distinguish progress from repeated error.
Learning may change an evaluation set, an approved knowledge base, a tool permission or an operating procedure. Model training is a separate, deliberate choice. Recording executions or adding memory does not by itself prove a performance gain.
Agentic iteration can also increase cost and propagate errors. We use explicit criteria and measured results to decide whether the additional complexity is worthwhile.
Within Verify: turn experience into a controlled change.
Evaluate → Reinforce → Verify / continuous improvement within Verify
After release, reviewed observations feed the next controlled change. Each change returns to the agreed evaluation criteria.
Observe
Capture the unexpected outcome, relevant inputs and system version with appropriate data access and retention.
Review
A responsible person investigates the cause and decides whether a change is needed.
Test
Make the finding a regression case. Test the proposed correction alongside legitimate tasks and other failure paths.
Approve
Review the evidence, remaining risks and rollback plan before releasing a versioned change.
Verify
Check the result in operation and retain the test, decision and procedure for the next review.
↶ Continue observing under the agreed review schedule
Illustrative example · payment agent
An invoice names a changed beneficiary, so the payment is held. The team investigates the discrepancy and adds it to the evaluation set. Any correction must preserve approval for the exact payment and pass the relevant tests before release. The case then becomes part of future regression testing.
Inspect the fictional workflow →Knowledge stays with the organization.
Centauro's delivery is designed to leave useful operating assets: versioned tests, approved procedures, control configurations and decision records that the client team can inspect and maintain. The engagement defines access, ownership and handover so those assets remain usable as people and systems change.
A report records what is known at a point in time. A maintained evaluation and review process gives the team a way to test the next change. That continuing capability is the value we aim to build, with maintenance responsibilities and review intervals agreed explicitly.
Governance makes improvement accountable.
Define who may change the system, which evidence supports approval and when to stop or roll back. Keep uncertainty and unresolved exceptions visible. These practices make learning reviewable and help the organization decide when greater autonomy is justified.
ISO/IEC 42001 provides a framework for establishing and continually improving an AI management system. We connect that management discipline to concrete controls, tests and operating decisions. Certification and measured agent performance answer different questions; neither should be inferred from the other.
Measure the learning curve you actually have.
Agree a baseline, a representative set of cases and the quality conditions that must hold. Track cost per successfully completed task, human review effort, error recurrence and recovery time across versions. Include model usage, investigation and maintenance in the cost.
Progress may be uneven. A necessary control can add cost while reducing exposure. We do not promise a fixed cost reduction with each doubling of executions. A test result remains conditional on its inputs, implementation and scope; wider autonomy needs additional evidence.
The ambition is durable capability: a team that retains what it learns and can demonstrate when its system is ready to do more.