Operate: quality & scale
Observability of the full runtime
Content-blind tracing of every run: structure, timing, and spend recorded, deal data never. Dashboards per agent and per tenant cover cost against each run's declared budget, latency per tool call and chain, gate metrics in both failure directions, and eval score trends. The audit trail is append-only, and every generated deliverable is logged with its template version, so outputs stay as traceable as the values inside them.
Token optimization and performance
The operational loop tunes what the traces reveal: model routing mix within plane cost ceilings, compaction policy, caching, parallel tool calls wherever the dependency graph allows, and effort scaling so simple jobs never spend like complex ones. Every tuning lands as a versioned configuration change through CI; silent drift is a defect. Clean-instance comparison is the test bench before any tuning ships.
Durability and the AI-debt checklist
| Rule | Why it prevents debt |
|---|---|
| Generated views, never hand-edited | Documents and configs cannot drift from their sources |
| Registry-only deploys | No shadow agents; nothing unregistered can obtain credentials |
| Build-twice rule | Shared assets are promoted on the second occurrence, never speculatively |
| Deprecate, never delete | History stays reconstructable; rollbacks stay possible |
| Versioned everything | Ontology pins, model pins, config releases: reproducibility survives improvement |
| Two sealed loops | Operational tuning and knowledge promotion never contaminate each other |
Scale-out
Growth follows the SSRA arithmetic rather than a rewrite: a new use case adds one skill file, a new client adds one binding and one generated constitution, a new data source adds one south server registered in the library. The playbook re-runs per engagement, and every run makes the next one faster because overlays, connectors, and agents compound in the platform.