Reported operating use · Software product quality assurance

Cambri’s reported agent-assisted QA with a human verdict

This is an attributed internal QA account published 9 November, an explicit older research fallback. Evidence artifacts and control enforcement were not independently inspected; the separate proposed rehearsal remains unexecuted.

Cambri’s reported agent-assisted QA with a human verdict

Cambri internal QA team, reported by CTO Roman Nikolaev

Author-reported internal operating QA, published 9 November 2026; an explicit older seven-day research fallback, not a just-published account or externally verified deployment.

A published third-party example, not a Nuvepro customer case or evidence of a partnership.

The workflow at a glance

Cambri reports agent-assisted testing with a human verdict

Illustrative summary of the published account

  1. QA engineer and Claude Code toolkit

    Scan waiting tickets and sort the queue

    The engineer starts a morning scan; an agent maps change impact, runs existing tests and sorts tickets waiting for QA.

  2. Next
    QA engineer

    Choose the testing scope

    The agent suggests a session scope and the person chooses what to test.

  3. Next
    Testing agent

    Challenge claims and explore edge cases

    The toolkit breaks criteria into claims and names falsifying evidence. It checks existing tests or code before remaining browser work, then explores edge cases.

  4. Next
    QA engineer and local evidence board

    Review evidence and make the final call

    QA reviews scenarios and evidence, answers questions and requests more cases. The agent never marks a ticket Done and never files a bug without approval.

Internal QA reported by Cambri’s CTO on November 9, 2026, an explicit seven-day fallback. Artifacts, runtime and permission enforcement were not inspected. The execution-share estimate was unmeasured; fully automatic low-risk testing had not happened and few tickets were classified low risk. Requirement checks were abandoned, defect causes remained unclear, and parallel agents or a future quality harness were prospective. See the source account and limitations below.

What the source reports

Cambri CTO Roman Nikolaev reports that QA engineer Javier built a Claude Code toolkit. A morning scan maps ticket impact, runs existing tests and sorts the queue; a person chooses the work. The agent breaks acceptance criteria into falsifiable claims, checks existing tests or code before browser work and explores edge cases. QA reviews the evidence and questions, requests more scenarios and retains the final verdict. The author says the agent cannot mark a ticket Done and cannot file a bug without approval.

Reported technology: Javier’s Claude Code QA toolkit with human scope and verdict. Scope and outcomes are source-reported.

What the enterprise environment would need

The following is a proposed project design for an enterprise doing similar work. Published details are marked; the remaining systems, data and scenarios would be agreed for your project.

Proposed practice environment

Rehearse ticket testing against independent fault fixtures

Illustrative project design to scope with your enterprise.

  1. 1 · Set up the environment

    Synthetic tickets and a frozen test workspace

    Use approved synthetic tickets, explicit acceptance criteria, a frozen code/app version and separately prepared expected outcomes and seeded-fault fixtures. Isolate the environment from production, real customer data and ticket or bug-writing endpoints.

  2. 2 · Preparation and agent contributions

    Agent prepares a falsifiable test evidence pack

    Propose scope for human selection, map claims to disconfirming cases, inspect tests/code and exercise the disposable app. Record scenarios, versioned screenshots/logs, unresolved questions and draft findings without setting ticket status.

  3. 3 · Independent checks

    An independent verifier replays reference cases

    A verifier separate from the generating agent compares evidence with independently authored outcomes and known-fault fixtures. Reproduce failures and check expected non-failures, boundary combinations, stale-version detection and evidence completeness; agent agreement is not verification.

  4. 4 · People decide acceptance

    QA owner accepts the exact rehearsal verdict

    The accountable QA owner accepts or rejects the exact independently checked scenario set, evidence pack and unresolved issues. This is a bounded rehearsal disposition; no live ticket-Done, bug filing, code merge or production release is authorized.

Only within the accepted practice scope
Practice output

A reviewable evidence package

A versioned rehearsal evidence pack with selected scope, falsifying scenarios, independent reference results, unresolved issues and QA-owner disposition.

If a check fails → withhold acceptance.

A missed seeded fault, false pass/failure, stale app version, missing evidence, unauthorized write request or owner rejection withholds acceptance. Repair scope, tests and evidence, repeat independent fixture verification, then return the exact revised pack to the QA owner.

Failure or owner rejection returns to review

  1. Hold
  2. → 2 · Revise work
  3. → 3 · Rerun independent checks
  4. → 4 · Owner review

Operating boundary. Designed checks, not executed. Synthetic tickets and disposable app only; no live ticket status, bug filing, merge, customer action or deployment authority. QA and product owners must confirm reference behavior and acceptance rules.

Proposed, not an executed Nuvepro project or a description of the named company’s implementation. Owners agree the data, rules, responsibilities and acceptance criteria for their enterprise.
Synthetic tickets and a frozen test application
Proposed: a QA owner supplies approved synthetic tickets, acceptance criteria, a versioned application snapshot and test data in an isolated workspace. A separate specialist prepares expected outcomes and planted defects before execution; freeze test, code and browser versions.
Human-selected claim and evidence plan
Proposed: an agent maps impact, summarizes existing test results and proposes a sorted queue. A person selects scope. For every acceptance claim, draft a falsifier, use test or code evidence first and identify which unresolved claims require browser checks and edge scenarios.
Independent result and coverage checks
Proposed: a QA reviewer separate from the generating agent compares observed results with the frozen expected outcomes, replays selected scenarios, checks planted-defect detection and inspects raw logs or captures. Passing code tests do not substitute for required browser behavior, and generated evidence cannot verify itself.
QA-owner acceptance and repair
Proposed: the accountable QA owner accepts or rejects the exact independently checked rehearsal evidence package, including untested claims and unresolved questions. A changed build, missing scenario, failed check or rejection withdraws acceptance; repair the test or evidence, repeat independent checks and return the new version to the owner. No live ticket update, bug filing, merge or release tool is enabled.

A proposed simulation

An illustrative practice design derived from this work, not an executed assessment.

Proposed: prepare a claim-by-claim rehearsal QA evidence package for human-selected synthetic tickets, with explicit falsifiers, independently checked outcomes and a human verdict.

Review gate: The QA owner accepts only the exact rehearsal package whose selected claims, scenarios, observed results and coverage passed independent checks against frozen references. Acceptance grants no ticket completion, bug filing, merge or release authority.

Challenge the result: A browser regression remains despite passing existing tests, or the build changes after capture. Withhold acceptance, repair and replay the affected scenarios, repeat independent result and coverage checks and return the revised package to the QA owner.

What needs checking

  • Source boundary: this is the CTO’s internal operating account. The actual toolkit, code, local board, screenshots and videos were not inspected; reported human verdict and approval rules are not independently audited enforcement.
  • Proposed: test a true acceptance claim, a misleading passing unit test, a planted browser regression, an empty or maximum input, interacting settings and a mid-flow refresh against separately prepared expected outcomes. Designed, not executed.
  • Proposed: change the build after evidence collection or reject an incomplete packet. Withdraw acceptance, repair and replay affected scenarios, repeat independent result/coverage checks and return the exact revised evidence package to the QA owner. Designed, not executed.

Review environment fidelity, human capability and agent performance separately. The enterprise’s accountable owners decide acceptance of the shared workflow and any operating use.

How an enterprise could build similar work

A proposed build order for an enterprise doing similar work, written by Nuvepro from the published account and our interpretation of the environment and checks above. It is not how Cambri internal QA team, reported by CTO Roman Nikolaev built or operates this work.

  1. Freeze synthetic tickets and independent expectationsProposed: prepare approved tickets, criteria, application/test versions and separately authored expected outcomes and planted defects.
  2. Let a person select the claim-checking scopeProposed: an agent suggests queue order and falsifiers, using existing tests or code before unresolved browser checks.
  3. Collect and independently inspect evidenceProposed: a separate QA reviewer compares logs, captures and replayed scenarios with frozen expectations and required coverage.
  4. Repair and repeat affected scenariosProposed: withdraw acceptance for changed builds, failed checks or missing coverage; repair evidence or tests, then independently recheck.
  5. QA owner accepts the exact rehearsal verdict packageProposed: retain untested claims and unresolved questions; accept or reject the checked version without operational ticket or release authority.

Your systems, data, rules and accountable owners decide the real project. Agree acceptance criteria, escalation owners and operating responsibility with the people accountable for the work before anything runs on live records. Treat the first run as a rehearsal on test records: compare the draft output with an independently prepared answer, and record what the people reviewing it change.

Questions about this example

Who chooses the work and makes the final QA decision?
The source says a person selects the scope and QA reviews evidence, questions and extra scenarios before the final verdict. The agent cannot mark Done and cannot file a bug without approval; enforcement was not independently audited.
Does Cambri report fully automated low-risk QA?
No. The author says that has not happened, the classifier rarely calls work low-risk and shift-left requirement testing was abandoned. Parallel QA agents and harness engineering are future directions.
Were the execution estimates independently measured?
No. The author explicitly calls them a team estimate from a week of sessions. The toolkit, code, board and screenshots/videos were not inspected, and no numerical impact or defect-cause conclusion is claimed here.
What would the proposed rehearsal add?
Approved synthetic tickets and a frozen application, separately prepared expected outcomes and planted defects, independent result/coverage checks and exact-version QA-owner acceptance. No live ticket changes, bug filing, merge or release are enabled.

Shape this around your enterprise

Bring a similar workflow and the people responsible for it. Together, define the records and systems to use, the result your team will build or configure, and how its owners will check it. Nuvepro scopes the simulation, environment and preparation. Agree any existing-agent evaluation, implementation assistance and production responsibilities separately.

Sources and limitations

Curation review: . This is the date this source example was reviewed, separate from source publication dates.

  • The complete public written article was read; the linked earlier article and actual board, code, screenshots, videos and sessions were not inspected or run. No numerical impact is independently verified.
  • The execution-share figure is a team estimate from a week of sessions, not a measurement, and is not reproduced here. The author says fully automated low-risk testing has not happened and the classifier rarely marks work low-risk.
  • Shift-left requirement testing was abandoned under pressure. The author does not know how increased defect findings split between better detection and earlier code quality; no causal improvement is established.
  • Parallel QA agents and quality-harness engineering are prospective directions, not reported current operation. This 9 November account is an explicit older fallback; it is not FDE delivery, a Nuvepro customer or a partnership.
  • The separate proposed rehearsal and business checks are designed and unexecuted. QA owners must confirm expected results and acceptance rules; practice acceptance grants no live ticket, bug-filing, merge or release authority.
  1. We handed testing to AI. Here is what stayed human.Source publication: 2026-11-09

More software engineering examples

All 16 software engineering examples

Explore more enterprise work