Business to outcome, end to end — running live. Work enters at the top, becomes a map, then a business case, then a workforce design. Agents fan out in parallel and rejoin at evaluation. Nothing reaches production without passing the gate.
Enrichment, multi-source research and eval runs fan out concurrently. Every join carries an explicit reducer — the default is last-write-wins, which produces a wrong answer rather than an error, and that is far worse than a crash because nothing alerts.
You cannot model the economics of a process that has not been mapped. Baseline before deployment. Never two nodes writing the same record. Never a client-visible message sequence out of order.
Six of these owners are internal agents, each pinned to an autonomy level that never drifts upward without an explicit decision. Anything touching money, contracts or an outbound client-facing send is capped at draft-for-human-approval. Permanently.
| Stage | What happens | Owner |
|---|---|---|
| Business | The work as it exists today | Client |
| Process mapping | Steps, actors, systems, handoffs, failure modes | Cartographer |
| Business case | Cost, payback, sensitivity — and the verdict | Abacus |
| Workforce design | Which agents, what boundaries, what escalates | Architecture review |
| Agents (parallel) | Catalogue templates configured to the map | Wright |
| Evaluation | Golden sets, regression gate, deployment block | Assay |
| Deployment | Staging on synthetic data, then production | Delivery pod |
| Monitoring | Uptime, escalation, drift, cost per task | Sentinel |
| SkyBuild Proof | Measured outcome against the signed baseline | Ledger |
| Business outcome | What actually changed, with provenance | Client |