Build & promote

The path from an idea to something a whole department depends on: author, validate, evaluate, promote, and then keep watching. Every gate on that path is automated except the one that should not be — a human reviewer's decision.

Ways to start

The template pack

Sixteen first-party templates ship with the platform. Each bundles a starter evaluation set, so a template-born agent can clear the promotion floors with no evaluation authoring at all.

Twenty first-party skills ship alongside them — document, spreadsheet, deck and PDF toolkits, charting, data profiling and forecasting, research and citation helpers, digest and diffing utilities, writing helpers, and formatting tools for channels, calendars, and diagrams. They install into the org ring on request and are already scan-passing.

Validate

Validation runs as you edit and again on save. It checks the schema, slug uniqueness, every skill and MCP reference (both that it exists and that you can see it), version pins, and the model allowlist, and it lints the body for instructions that would undermine the platform's guardrails. Errors come back with field paths, so the editor can point at the line.

Version

Versions are immutable; the agent or skill row holds a pointer to the current one. That single design choice is what makes rollback safe and promotion meaningful.

• Rollback moves the pointer. Nothing is mutated, nothing is lost.
• Diff compares any two versions.
• Every version is scanned — five stages, individually reported. See Governance.
• Shared-ring versions are regression-tested. Publishing a new version re-runs the evaluation set; a failure reverts the pointer and notifies the owner, so the working version stays live.

Evaluate

An evaluation set is a list of cases, each with an input and one or more assertions. Cases run against the candidate version through the real execution path — the same executor that serves production traffic, not a test harness that might diverge from it.

Cases carry a category, and categories are what the promotion floors key on:

Evals live in the Playground, under the agent's Evals tab: add cases by hand or start from the starter set, run them against the current version, and read the verdict per case with its trace. The success rate, latency and cost per run are tracked per version, so a regression shows up as a step in the history.

Preflight: certify before you propose

Every gate the promotion runs, you can run yourself first — against your draft — in the Playground. The Certify panel reads the same checks the reviewer's queue will enforce: the definition validates, the security scan verdict, the eval coverage floors, and the eval run for this exact version. Because it reads those gates rather than inventing its own score, a green preflight and a passing promotion cannot disagree — you find out it is shippable before anyone is asked to review it.

Promote

Moving an agent from your personal ring to a team or the whole organization is a reviewed promotion. The gates run in order, and each one reports to the reviewer.

1. Propose

   The owner proposes their personal agent for a target ring.

2. Scan gate

   The version's security scan must have passed. A flagged scan blocks with the findings attached.

3. Floor gate

   The evaluation set must meet the floors: a minimum number of cases, and — for a write-capable agent — refusal, injection, and red-team coverage. Declared autonomy above the organization's floor is rejected here too.

4. Evaluation gate

   The eval run against the candidate version must be green, case by case.

5. Human review

   A reviewer reads the gate reports and the candidate definition, then approves, requests changes, or rejects. Requesting changes loops back to the owner without losing the thread.

6. Approve

   Approval copies the agent and version into the target ring and freezes its skill references to exact versions. The shared copy cannot drift while the author keeps editing their original — which stays on floating references.

7. Staged rollout

   Optionally approve to a pilot group first. The agent stays invisible to everyone else until it is rolled out.

Watch

Promotion is not the end of the story. Production runs are sampled continuously and graded on the cheap tier. If the rolling pass rate falls below threshold:

• The owner is alerted — once per period, not once per failure.
• A single demotion proposal is raised, riding the promotion pipeline in reverse.
• A reviewer approving it pauses the agent, which then refuses new runs at the execution path.

The builder dashboard plots the evaluation trend alongside success rate, latency, and cost, so a slow degradation is visible before it becomes an incident.

Share

The Skill Workshop

Agents notice repetition. When one does, it can call propose_skill to draft a playbook into the workshop. Proposals can be revised conversationally and then applied — which validates the draft and creates a real skill in the proposer's personal ring, triggering a scan — or rejected with a reason. A proposal is never a live skill, and there is no other path from one to the other.

The Agent Architect

A built-in meta-agent for authoring. It can read an existing definition, search the catalog for prior art, write a draft into a conversation-scoped buffer, and validate it through the same pipeline the editor uses. It deliberately cannot save — drafts become agents only when a human accepts them.