← nTra/How it works

From intent to a shipped product —
in four phases, on one loop.

Discover, Plan, Build, Ship. The same loop your team already runs. Agents handle decomposition, execution, and validation between human checkpoints — so every detail gets the time and craft a great product deserves.

The loop

One continuous cycle —
not a waterfall.

Every loop ends where it began — with new intent. Validation between phases keeps work honest; human gates between phases keep the team in control.

your ideashipped product01Discover02Plan03Build04Ship
Phase 01
Discover
Capture intent, structure context, surface the right questions.
Deep dive ↓
phase node human checkpoint active flow
Phase 01

Discover

capture

Conversations, briefs, prior decisions, and constraints flow in. nTra parses them into a structured context graph the rest of the system can reason about — and surfaces the ambiguities a senior PM would ask about before any work starts.

What happens
01
Parse intent
Briefs, threads, and discovery calls are decomposed into goals, constraints, and acceptance criteria.
02
Structure context
Dependencies, prior decisions, and codebase context link into a queryable graph.
03
Surface questions
Gaps, ambiguities, and risky assumptions are flagged back to the human for clarification.
Agents in this phase
OracleInquisitorAtlas
Human role
Founder · Product owner approve direction
Artifacts produced
Structured context graph
Open questions
Linked prior decisions
Acceptance criteria
Discover · visualized
Slack threadDiscovery callBrief.pdfCodebaseCONTEXT GRAPH? × 3
Phase 02

Plan

design

Architecture, design language, and an executable plan are generated together — so the system, the surface, and the sprint are coherent from the first line written.

What happens
01
Generate architecture
Services, data flow, contracts, and dependencies are scaffolded from the parsed model.
02
Design UI / UX / VX
Flows, layouts, tokens, and motion language are produced as a coherent surface — not just screens.
03
Create execution graph
Work decomposes into a directed graph of tasks with explicit dependencies and sprint slots.
Agents in this phase
ForgeAtlasArtisan · designMaestro
Human role
Architect · Tech lead · Design lead review and adjust
Artifacts produced
System architecture
UI / UX / VX kit
Execution graph
Sprint plan
Plan · visualized
SYSTEMWebAPIAuthDBSURFACEEXECUTION GRAPH11 tasks · 3 lanes · 2 gates
Phase 03

Build

execute

Design, frontend, backend, tests, and docs agents run in parallel — bounded by the execution graph. Every artifact is checked against the original intent before it leaves the agent that produced it.

What happens
01
Coordinate agents
Specialized agents pick up tasks from the execution graph and run in parallel lanes.
02
Validate in flight
Every artifact is checked against intent, constraints, and acceptance criteria as it is produced.
03
Reconcile
Interfaces between agents — types, contracts, design tokens — are kept consistent automatically.
Agents in this phase
Artisan · designArtisan · frontendArtisan · backendVerifierScribe
Human role
Engineering reviews PRs · merges on its own cadence
Artifacts produced
Branches and PRs
Generated UI + design tokens
Test suites
Docs and changelog
Build · visualized
designfrontendbackendtestsdocsVALIDATE
Phase 04

Ship

sync

Branches, PRs, tickets, changelogs, and deploys land in the tools your team already uses. Every decision and every artifact is logged — nothing about the team's workflow changes.

What happens
01
Open PRs
Branches and PRs land in the repo, attached to the originating intent and execution trace.
02
Sync tracker
Tickets, statuses, and assignments are written back to Jira, Linear, or your tool of choice.
03
Deploy on gate
Staging deploys automatically; production sits behind the human gate your org configures.
Agents in this phase
HeliosBridgeMaestro
Human role
Team ships on its release cadence
Artifacts produced
PRs in GitHub / GitLab
Tickets in Jira / Linear
Changelog entries
Deploys to staging / prod
Ship · visualized
nTragithub.com/acme/appPR #418 → mainlinear.app/acmeCSV-12 → in reviewchangelog · 2026-05-24+ CSV exportstaging.acme.comdeploy ok · 2m agoprod.acme.com⏸ awaiting gate
Walkthrough

One loop, end to end —
"users can export their data to CSV."

One intent. Sixteen steps. Watch a real loop traverse all four phases — from the command that starts it to the gate that waits on a human.

nTra · loop trace
$ nTra exec "users can export their data to CSV"
step 01 / 1601 · Discover
Human in the loop

Four checkpoints.
You stay in control.

nTra pauses at every gate. Nothing ships to production without a human saying so.

1
After Discover
Direction & scope
Founder · Product owner

You confirm the parsed intent matches what you actually want — and answer any clarifying questions nTra surfaced.

2
After Plan
Architecture, surface, plan
Architect · Tech lead · Design lead

Generated architecture, design system, and execution graph are reviewed — you can edit any of them before a single line is written.

3
After Build
Code & artifacts
Engineering

PRs land in your repo. You review and merge on your own cadence — nTra never pushes to main on its own.

4
After Ship
Production deploy
Release manager · ops

Staging deploys are automatic. Production sits behind whatever human gate your org configures.

Validation layer

Every handoff is
verified before it moves.

IntentPlan
Maestro + Inquisitor
Goals match constraints
Ambiguities resolved
Prior decisions respected
PlanBuild
Forge + Atlas
Architecture fits scope
Surface coherent with system
Execution graph is acyclic
BuildShip
Verifier + Sentinel
All acceptance criteria met
Type contracts consistent
Test coverage ≥ threshold
ShipLoop
Maestro
Artifacts match intent
Tracker up to date
Production gate respected

See the loop run on
your own idea.

Bring an intent. nTra brings the rest of the team.