Six stages. Two human sign-offs. Thirty-two specialist agents in between. A walkthrough of how a single ServiceNow story flows through the GeneWorks delivery loop — and why the loop never closes.
Every story — a catalog tweak, a business rule edit, a scoped app feature — runs through the same six stages: Requirement, Design, Task Planner, Execute, SN Deploy, Verify. The agents drive all of them. At two points, the loop pauses for a human: design sign-off before anything is built, and verification sign-off before it ships.
You talk to the Requirement agent. It turns whatever you bring — a Jira story, a CSV from a workshop, a Visio swim-lane, or plain English — into a structured, versioned Functional Requirements Spec, and asks clarifying questions where the requirement is thin. Too ambiguous to specify? It bounces back with a structured list of what's missing.
The Design agent produces a full Solution Design — data model, dependencies, ACL implications, integration shape, and the exact ServiceNow Fluent artifacts to build. It reads from the workspace's accumulated context, so it matches your team's conventions. Then your architect reviews and signs off — nothing builds until this happens.
The Planner agent breaks the approved design into a granular, ordered implementation plan — every table, role, group, system property, ACL, business rule, and test as its own task. You can regenerate or adjust the plan before a single line is built.
The Main agent builds the tasks in parallel waves, writing real ServiceNow Fluent code (React + TypeScript) you can read in the built-in IDE. ATF and functional tests are authored alongside the build — not after the fact. Each task is marked done or flagged, and every agent sees the others' work in the same workspace.
GeneWorks installs the app into your instance as a safe, non-destructive update. Any changes made on the instance since your last deploy are merged in first, so your work never overwrites theirs. Every step is checkpointed — the whole deploy is fully reversible. GeneWorks never deploys straight to production.
GeneWorks runs ATF (server + client) plus live-browser functional tests against the deployed app — on our own test runner, so you never license ServiceNow's ATF cloud runner. Failures get one-click auto-fix where the pattern is known. Then your team signs off on the evidence. No theater: if 73 of 100 stories completed clean, it says exactly that — with a handoff for the rest.
The gates aren't ceremony. They're the contract. Every change record points to a named human who approved the design and a named human who approved the evidence. Audit trail by construction, not by reconstruction.
The agents have specified what they will build, why, and how it touches the rest of the instance. Now a human says yes or no. No code has been written yet. Reversing course here costs minutes — reversing it after build costs days.
The app is deployed and verified — ATF plus live-browser functional tests, on our own runner (no ServiceNow ATF cloud-runner license). The agents produce an evidence package per story: what passed, what's flagged, what's blocked. A human signs off before anything goes to CAB. No story is "done" because an agent says so.
GeneWorks isn't one model wearing different hats. It's a roster of specialist agents — each scoped to a discipline, each with their own tooling, each with the prompts and patterns that produce reliable output for their artifact type. Gene is the orchestrator that routes work between them.
Cold-start automation is harder than mature automation. The first month, the agents are learning your instance — your idioms, your ACL patterns, your naming conventions, your integration map. By month three, the failure log and the design pattern library are doing real work. By month six, the agents read like the team.
| Phase | Window | Effort reduction | What's happening |
|---|---|---|---|
| Cold start | Week 1–4 | 25–35% | Agents learn the instance, team idioms, infrastructure map. |
| Context maturing | Month 2–3 | 40–50% | Failure log and design pattern library accumulate in workspace. |
| Compounding gains | Month 4–6 | 55–70% | Agents propose like the team. Reviewer flags drop. |
| Steady state | Month 6+ | ~65% | Plateau. The remaining work is genuinely hard. |