Hands-on evaluation
The clearest next step Allvue asked for: give a couple of their people hands-on access, both to evaluate internal use and to see where cortex fits in sellable products. This page proposes a structured evaluation rather than an open sandbox: three exercises, each proving one of the points the review questioned, with what the evaluator sees at each step.
Setup
| Item | What Allvue receives |
|---|---|
| An organisation on the platform | An organisation on the development environment with two evaluator accounts as organisation administrators; sign-in by password, or by single sign-on once the issuer is configured for Allvue |
| A project | An empty project paired with its own engine fork, and read access to a finished reference project so the evaluators can compare |
| An MCP access token | Minted in Settings, so an evaluator can also drive the project from an IDE or a headless client |
| This documentation and the product documentation | The pages here plus docs.cortex.mindshine.com |
Exercise 1: build an agentic workflow against a sample API
What it proves: how tools, workflows and agents are represented, and that an agent's reach is bounded.
- Upload a short brief as a requirement source: a process with one human approval and one automated step that reads a sample API. Watch the requirements appear with their citations.
- Approve the build preview. One card is filed; the fleet builds the endpoint that reads the API, the workflow with its steps and the agent actor's manifest, and attaches evidence.
- Open the workflow canvas, start a run, act on the human step from My Work, and read the actor's output on the run.
- Open the Agents page: the actor's tools are exactly the endpoints its manifest names; try to add a tool that does not exist and watch the manifest be refused.
Exercise 2: create or modify a reusable tool
What it proves: an endpoint is one thing used three ways, and every change is versioned.
- File a change card: "the API tool must also return the record's owner".
- When the card completes, open the endpoint in the studio: read the version history and the diff.
- Call the endpoint as a tool from an MCP client with the access token, then open the application screen that uses it and the test that asserts on it. Same endpoint, three callers.
- Restore the previous version from the history and watch the test go red on the next run, then restore again.
Exercise 3: change a requirement, regenerate, test, inspect the code and history
What it proves: developer control, contract governance and delivery.
- Upload a change request as a new requirement-source version. Observe that the existing requirements are not regenerated; only the affected rows are reworked, as a delta.
- Before filing the card, add an
httpcontract test on the existing API that asserts the current response shape. - File the card. If the change breaks the contract, the card fails on that test and says so; otherwise it completes with evidence.
- Open the App Services page, edit one file of the service by hand, save and rebuild. Then file a second small card and confirm the hand edit survives.
- Export the project (Deployments, Download for local run, Customer archive). Open the archive: the app service's Dockerfile and tree, the compose file, the runbook, the tests. Start it with one compose command on a laptop and run the tests inside it.
What each exercise answers
| Review question | Exercise |
|---|---|
| What exactly gets deployed | 3, step 5 |
| How workflows invoke tools; how agents are represented | 1 |
| Whether cortex is required after deployment | 3, step 5 |
| Developer override and source of truth | 3, step 4 |
| API contract governance | 3, steps 2 and 3 |
| Isolation and tenancy | 1, step 4, and the reference project's data policies |
The architecture validation session
The evaluation runs alongside one working session with Allvue's architects, built around the questions rather than a demo: deployment architecture with a real export on screen; identity and security against Allvue's own OIDC setup, walking through Security and identity with the gaps named; developer control (generate, edit by hand, version, change again); API and version governance with a contract test turning red on purpose; multi-tenancy on the reference project's policies. The pages here are the pre-read.