Software delivery has
been reinvented.
Are your teams ready?
The AI Product Studio replaces the traditional sprint cycle with a continuous, agent-assisted delivery loop — where AI agents and humans collaborate at every phase of the delivery lifecycle. Ideas move from intent to deployed feature in hours, not sprints. Governance, quality and security are built in, not bolted on.
The AI Delivery Lifecycle
Traditional Agile compresses delivery into sprints. The AI Delivery Lifecycle (AIDLC) compresses it further still — replacing human execution with agent execution at every repeatable step, while keeping humans in control of intent, decisions, and governance.
It's not a tool upgrade. It's a team composition change. Your Product Owner still sets direction. Your Engineering Lead still owns architecture. Your Compliance team still signs off. But at every level, an AI Agent works alongside them — drafting, coding, reviewing, testing, and shipping at machine speed.
Timelines collapse. Quality rises. Audit coverage becomes automatic. And because every agent action is policy-bounded and audit-logged, you get faster delivery without giving up governance.
Mindset Shift
Stop thinking "AI assists humans." Start thinking "humans govern agents." Humans own intent and decisions; agents own execution.
Compressed Timelines
When agents code, review, test and document in parallel, sprints that took two weeks take two hours. Your pipeline never sleeps.
Built-in Governance
Every agent action is policy-bounded, cost-tracked and audit-logged. Compliance is a by-product of how the system works, not a post-release checklist.
Scale Without Headcount
One hybrid team of 5 humans and 8 agents can deliver what a traditional team of 20 can — with better traceability and no bus-factor risk.
Your Hybrid Team — Humans and Agents, Every Role
Every role in the lifecycle has a human counterpart and an AI agent partner. The human sets intent, judges quality, and owns accountability. The agent handles the volume, speed, and consistency humans can't match alone.
Choose Your Trust Level
Not every team is ready to hand full execution to agents — and that's fine. The studio supports three operating modes. You set the trust level; agents operate within it. You can graduate between levels as your confidence grows. Supervised is on every plan; Semi-Autonomous and Fully Autonomous are part of the Business and Enterprise plans.
Supervised
You stay in the cockpit. AI is your co-pilot.
Every agent action is proposed and reviewed by a human before execution. Ideal for regulated environments, early AI adoption, or any domain where a human expert must validate before every step proceeds.
- Human approves every agent plan, step by step
- All generated code and artefacts reviewed before commit
- Manual release sign-off at every stage
- Full immutable audit trail
Semi-Autonomous
Agents do the work. You make the decisions.
Agents apply what they write and accept what they suggest, within the cost and risk thresholds you define, and runs auto-pass unless they are flagged. A person still approves each work package and decides what ships. Business and Enterprise plans.
- Hard cost, risk and topic guardrails enforced at runtime
- Automated security, quality and compliance gates
- Auto Pilot available — the Studio fills itself in and you review the result
- Agent pages, recommendations and Product Manual revisions are applied as they are raised, with the audit trail of an automatic decision
- A person approves each work package and decides what ships
Fully Autonomous
The studio drives. You steer.
Fully Autonomous runs the whole loop, from discovery to a pull request, with no manual input, under the owner's standing approval. People see everything it does on the Autonomous loop page and steer rather than drive. Business and Enterprise plans.
- Discovery → pull request, with no manual step in between
- Chosen by the studio owner: a standing approval, recorded with who and when
- Started with a spend ceiling; anyone in the studio can pause it, stop it or reorder its work
- Full visibility of what the loop is doing, and the means to steer it
- Merging the pull request stays with GitHub or a person
What each level gates, phase by phase
Taken from the product's trust-gate rules — your studio's Trust Settings page shows this same table, generated live from the rules that actually run.
| Phase | 🧑✈️ Supervised | 🤝 Semi-Autonomous | ⚡ Fully Autonomous |
|---|---|---|---|
| 💡 Intent | 🧑 Human gate | ⚡ Auto-pass* | ⚡ Auto-pass* |
| 📚 Context | 🧑 Human gate | ⚡ Auto-pass* | ⚡ Auto-pass* |
| 🎯 Plan | 🧑 Human gate | ⚡ Auto-pass* | ⚡ Auto-pass* |
| ⚙️ Code | 🧑 Human gate | ⚡ Auto-pass* | ⚡ Auto-pass* |
| 🎭 Review | 🧑 Human gate | ⚡ Auto-pass* | ⚡ Auto-pass* |
| ✅ Verify | 🧑 Human gate | ⚡ Auto-pass* | ⚡ Auto-pass* |
| 🚀 Release | 🧑 Human gate | ⚡ Auto-pass* | ⚡ Auto-pass* |
| 📡 Telemetry | 🧑 Human gate | ⚡ Auto-pass* | ⚡ Auto-pass* |
* Auto-pass applies to unflagged runs only. In every mode, runs flagged with security findings or rejections wait for human review, runs whose projected cost exceeds the agent's approval threshold are held for approval, and releases require the two-key sign-off (automation + a named human). In Supervised and Semi-Autonomous studios a person approves each work package; in Fully Autonomous, once it arrives, the owner's standing approval does.
The AIDLC — Step by Step
The same phases, three different trust tracks. Common steps span all three columns. Steps that differ per trust level show what each mode does differently.
Intent Capture
All Trust LevelsDrop any idea, problem, bug, or opportunity — voice note, screenshot, free-form text, or imported GitHub Issue. The Intent Classifier agent enriches and classifies it automatically. No structure required; agents handle the rest.
- Natural language, voice, image or structured import
- AI-assisted triage, classification and context enrichment
- Direct dispatch to GitHub Issues for agentic coding sessions
Context Library
All Trust LevelsFeed your agents with the right information — API contracts, architecture decisions, security constraints, persona definitions, and domain knowledge. Better context produces safer, more accurate, more consistent outputs. Every agent run in the studio inherits the library automatically.
- Specs, docs, rules, personas, API contracts and ADRs
- Version-tracked and audit-logged on every change
- Automatically injected into every agent run in this studio
- Ideas Board companion: a canvas-style desktop for raw, unstructured ideas (PDFs, voice notes, pasted text, screenshots) that an AI conversion task summarises into clean Markdown. Raw uploads are kept securely for up to 180 days; AI summaries are kept until you remove them.
Agent Selection & Planning
Human manually selects and configures each agent for the task. Reviews the proposed execution plan step-by-step and approves before any action begins.
AI self-selects from your approved agent pool within cost and risk caps, and the plan's pages are applied as they arrive. A person approves each work package before it goes to a coding agent.
The loop plans with no manual input, and a work package is approved under the owner's standing approval only once every reviewer agent has passed it.
Code & Artefact Generation
Code Agent drafts all implementation. Human reviews every file in the diff before it is committed. Nothing lands without explicit approval.
Code Agent generates, self-reviews and passes security checks within policy guardrails. Rework the reviewer agents ask for goes back to the coding agent as soon as it is recorded. Human reviews flagged items — not every file.
From the approved package to the pull request with no manual step, all within policy guardrails, while people watch the work and steer it.
Review
The Definition of Done's agents read every pull request and say what falls short. The Review feed gathers their rework, their questions and stale evidence across every package, and a person answers, reassesses or approves — each decision recorded with a name and a time.
The reviewer agents' verdicts decide whether the loop carries on; people see what the automated governance gates flag — security issues, policy violations, cost overruns — and merging stays with GitHub or a person.
Verify — Quality & Security Gates
All Trust LevelsVerify collates what the Definition of Done found across every package in delivery — by concern (security, accessibility, performance, compliance, tests, documentation, code quality), by pull request next to your own CI checks, and as a ledger of every finding.
- Your GitHub check runs and commit statuses beside each merge hold
- Configurable Definition of Done — a package cannot complete, or ship, until it passes
- Every finding kept, with the pull request head it was found on
Release
Releases are planned in work packages whose Definition of Done has passed, grouped by release phase and workstream, approved on two keys — the server's readiness check and a named person — and marked released only once every pull request has merged. Release comms for customers, internal teams and auditors are drafted from the release record and approved before they go out.
Manual human sign-off required at every stage — staging, pre-prod and production. AI-generated release notes are produced but a named human must approve before deploy.
Automated staging and pre-prod deployment. Human approval gate for production only. Release notes generated automatically; a named person reviews, signs off and decides what ships.
The loop takes the work as far as a pull request; merging and release stay with GitHub or a person, with approval records written to the immutable audit log.
Telemetry & Learning
All Trust LevelsEvery release names the key results it expects to move. A week after release, and again at each horizon, the Telemetry Analyst judges each one from the readings that actually came in — on track, behind, regressed, or inconclusive when there is nothing to judge. A release that falls behind raises a course correction for a person to accept.
- Metrics and signals — bugs, regressions, incidents, tickets, customer feedback — kept as evidence
- Release-to-outcome traceability, and intents that name the release they correct
- Effectiveness dashboard — AI ROI, agent accuracy, cost per feature
From Blank Studio to Shipped Code — Without Leaving the Flow
AI Product Studio guides you through every step: one-time setup in minutes, then a guided loop from idea to deployed feature with hard approval gates at the moments that matter.
🛠️ One-Time Studio Setup
A getting-started banner names your next setup step from every page, until the studio is ready to dispatch its first work package.
- ① Set your Trust Level — Supervised, or Semi-Autonomous or Fully Autonomous on Business and Enterprise
- ② Configure your AI Provider — paste an API key, select models
- ③ Hire your Agent Roster — enable the roles you need
- ④ Connect Code Repositories — the Codebase Analyser scans them and suggests intents & context
- ⑤ Seed the Context Library — architecture notes, NFRs, threat model
- ⑥ Capture Vision & Outcomes — every intent traces back to a strategic goal
🔄 The Continuous Loop — Per Intent
Voice notes, images, files, or free text — the Intent Classifier triages them instantly.
Enrich and edit the intent, link it to outcomes, run the Architecture Advisor. Say which intents come first: the Pad’s Map lays them out in that order, as advice, never a block.
On the intent's page, AI drafts the brief, proposes workstreams and a Work Package under each, then fills every package with estimated work items and BDD acceptance scenarios — each package carrying your Definition of Ready and Done. 🔒 A named human approves every package before dispatch.
The approved Work Package goes to a coding runner — the Studio's own system runner, GitHub Copilot or Claude — and the change arrives as a pull request. The package's Delivery stage and the Run Board show every run, its branch and its pull request, and link each back to what started it.
The package's Definition of Done agents judge the code change itself, pin their evidence to the commit they read, and send anything that falls short back to the coding agent — or have the system runner correct it on the same branch. 🔒 A named human still decides.
A commit status on the pull request stays pending until the Definition of Done passes. Require it in branch protection and nothing merges early; a new commit puts the hold back until it is assessed.
Release Publisher packages artefacts, generates structured release notes, triggers deployment.
Telemetry Analyst closes the loop — correlates runtime signals back to the originating intent.
🗳️ Agent Recommendations — at every phase
Whenever an agent surfaces a risk, a new context item, a policy suggestion, or an intent derived from a codebase scan, it goes to the Recommendation Queue — not applied automatically. You approve, reject, or edit each one before it becomes live data.
Auto Pilot
Once a Studio is running in Semi-Autonomous mode, the repetitive half of the loop can run itself. Auto Pilot keeps generating until the agent has nothing new to add, then hands you the result to review.
Take it from here
From one intent, Auto Pilot writes the brief, breaks it into workstreams, creates a work package under each and fills every package with work items, then stops, leaving each package as a draft for you. It never moves you: the panel says where it is working, with a link to watch that page fill in, and you carry on with something else when you would rather not.
Or one surface at a time
The same loop runs on its own wherever a page produces a list: context, risks, policies and then intents from an Ideas Board, each intent naming the ones to implement first and carrying the board’s notes; outcomes and their key results on Vision; glossary, personas, architecture, NFRs and threat model in the Context Library.
Stop at any point
The panel follows you across pages, so the stop button is always to hand. Pause keeps what has been produced so far; stop lets the current step finish rather than orphaning its work. A run that was stopped, paused or cut off by a closed browser can be resumed from where it got to, without paying for the finished stages twice.
It knows when to stop
A run ends when a pass produces nothing new, when a stage reaches its item cap, or after repeated failures. A run budget also stops and asks whether to carry on — or you can tell it not to ask. Every generation is told what already exists, so a longer run does not fill up with rewordings of what you already have.
Auto Pilot stops at draft work packages. Sending one for review, approving it and sending it to an agent remain human decisions, in Supervised and Semi-Autonomous studios alike.
Governance & Control
AI Product Studio is built for regulated environments. Governance is part of the workflow, not an afterthought.
DPIA & Model Risk
AI-assisted Data Protection Impact Assessments and Model Risk Assessments built into the platform. Track completeness, get AI suggestions for each field, export to Markdown for your DPO.
Full Audit Trail
Every agent run, human decision, cost, approval, and rejection is logged with timestamps, actor IDs, and full context. Append-only — nothing can be deleted. Configurable 12-month retention.
Policy Guardrails
Define explicit policies for AI usage — off-limits topics, data residency rules, cost limits, and escalation triggers. Agents operate inside your guardrails, not around them.
Human-in-the-Loop Gates
Every stage has a configurable approval gate. Set cost thresholds, risk levels, or specific action types that require a named human approver before proceeding.
Merge Held Until Done
Every pull request a coding agent opens carries a Definition of Done status. Evidence is pinned to the commit it judged, a new push marks it stale, and branch protection keeps the merge closed until the gate has passed.
An Artefact Passport per Work Package
Every check, decision and exception on a work package or a release is recorded in an append-only evidence ledger as it happens, and shown as one page: its Passport. Checklist items name the controls they evidence, and an exception records who granted it, why and until when.
Your Agents, Your Rules
Import the agents your repository already defines — Copilot custom agents and Claude sub-agents — and name them on checklist items. They review under their own instructions; what they must verify stays fixed, and every change to an agent is audited.
Workspace Isolation
Each studio is fully isolated. No context, run data, or artefacts cross tenant boundaries. Ringfencing, IP restriction, MFA, and SSO are configurable per workspace.
Cost Transparency
Every agent token is counted and costed. Per-run USD cost is visible in real time. Set hard caps per agent, per run, or per studio. No surprise bills.
A plan for every agent-led team
One subscription covers the AI Product Studio and the Product Workspace, priced per person: someone who works in both counts once. Every plan uses your own AI provider keys, and you upgrade without recreating your Studio or moving data.
One developer
Start with the governed default agent roster.
- 1 developer
- Default agents
- BYOK AI providers
A development team
Build a tailored fleet, for as many developers as you buy seats for.
- Custom agents
- Marketplace agent recruitment
- BYOK AI providers
Across the organisation
Govern agents and integrations across every studio.
- Organisation-wide agent management
- Organisation-wide integration management
- Organisation guardrail policies
- BYOK AI providers
Who Uses AI Product Studio
Product Owners & Managers
Capture ideas faster, get AI-scoped briefs in seconds, and stay in control of what ships — without managing every code review.
Engineering Leads
Set cost guardrails, review agent plans before execution, and let AI handle the repetitive scaffolding work while engineers focus on hard problems.
Compliance & Risk Teams
DPIA templates, model risk assessments, append-only audit logs, and policy guardrails give you the evidence trail needed for AI governance frameworks.
Startup & Scale-up Teams
Move at the speed of a team twice your size. Intent to deployed feature in hours, not sprints — with the audit trail you'll need when you scale.
Ready to run your first AI sprint?
Create your studio in 60 seconds — free plan included. No credit card required.