Skip to content
AI FlywheelStudio
Join early access
DO · Opens Feb 2027 · Early access

The AI agent builderAgentic Studio

Agentic Studio is the AI agent builder of AI Flywheel: build agents that act, with the system that watches them.

Join early access Sign in
Close-up of code on a dark monitor, the kind of spec an AI agent builder works from
agentspec.yamlIllustrative
agent: intake
owner: a named human
risk_tier: 2
autonomy: L3  # proposes; a person approves
tools: [mcp://docs.extract]
evals: golden/intake  >= 0.90
kill_switch: tested

What you build

One agent. One bounded workflow.

Start with the wedge chosen at G1, then expand by reuse.

01

One bounded workflow

An agent with a named owner, a risk tier and autonomy limits.

02

Eval suite

Graded test cases that grow with every spec.

03

Gate log

Six signed decisions, G1 to G6, with linked evidence.

04

Tested kill switch

Proved to stop the agent before go-live.

ADLC

Ten phases, six gates.

A gate is a signed decision with linked evidence, not a meeting.

  1. 00

    Discover

  2. 01G1

    Define

  3. 02G2

    Design

  4. 03G3

    Specify

  5. 04G4

    Prototype

  6. 05

    Build

  7. 06G5

    Verify

  8. 07G6

    Govern and release

  9. 08

    Operate

  10. 09

    Learn and evolve

The agent development life cycle (ADLC), phase by phase

G1 · after Define

Thesis approved

Is this worth building, and safe to build?

G2 · after Design

Architecture approved

Will this design meet the targets, risks and rules?

G3 · after Specify

Definition of ready

Can agents build this without guessing?

G4 · after Prototype

Experience sign-off

Do users understand it and want it?

G5 · after Verify

Quality gate

Does it work, safely, at load?

G6 · after Govern & release

Governance and go-live

May it go live, and can we stop it?

Inside the studio

The platform under every project.

Designed in from day one. See the platform architecture and how we handle AI agent governance.

Model Mesh

Model-agnostic routing by task, with residency rules. Bring your own key.

Connector Fabric

An MCP gateway, a connector registry and a token vault.

Eval Service

Quality gates, a replay bench and an evidence pack per release.

Policy and approvals

Autonomy limits enforced at run time; an approvals inbox for people.

Trust Ledger

A hash-chained record of who decided what, and on what evidence.

Kill-switch hierarchy

Run → Agent → System → Tenant → Global. The narrowest stop is always there.

Artefacts

Proof you can hand over.

Every project leaves four records behind.

docs/releases/gates.md

Gate log

Six signed decisions, each with linked evidence.

  • G5 quality gate pass
  • signed eval lead
  • evidence #evals #security
evals/intake.yaml

Eval suite

Graded cases that grow with every spec.

  • cases golden/intake
  • threshold 0.90
  • critical 0
registry/agents.csv

Register row

One agent, one named human owner.

  • agent intake
  • owner named human
  • autonomy L3 · tier 2
ops/kill-switch.md

Tested kill switch

Proved to stop the agent before go-live.

  • level agent
  • drill passed
  • restore human only

Autonomy dial

Optimal autonomy, not maximum.

Slide the dial. Each level needs its controls in place first.

2026 sweet spot

L3Delegate

Proposes actions; a person approves.

When in doubt, set the dial one notch lower. Raise it as trust is earned.

Controls required first

At L3: 4 controls required.

  • Required: Normal review · L1
  • Required: A person reads every output · L2
  • Required: Tests, evals and AI review before human review · L3
  • Required: Scoped permissions · L3
  • Not yet required: Sandbox and outbound limits · L4
  • Not yet required: Agent identity · L4
  • Not yet required: Audit log · L4
  • Not yet required: Sampled review · L4
  • Not yet required: Kill switch · L4
  • Not yet required: A long eval record · L5
  • Not yet required: Instant rollback · L5
  • Not yet required: Regulator-ready records · L5
Two learners reviewing code together on a laptop, as an Agentic AI Engineer cohort team does

Learn it in a cohort

Agentic AI Engineer

View the programme

Quick answers

Questions, answered.

The Agentic Lab on campus Agentic Studio for enterprise teams Join Agentic Studio early access

What do you build in Agentic Studio?

One bounded AI agent for one real workflow, chosen at G1 and expanded by reuse. Each agent has a named human owner, a risk tier and autonomy limits, and reaches tools through the MCP gateway. Agentic Studio is the AI agent builder that opens first, in February 2027.

Is Agentic Studio model-agnostic?

Yes. You bring your own key, and the Model Mesh routes each task to a suitable model under the tenant's residency rules, with fallbacks and guard models. Tools connect through an MCP gateway with a connector registry and a token vault, so no agent holds long-lived credentials.

What are the six autonomy levels L0 to L5?

L0 Manual, L1 Assist, L2 Pair, L3 Delegate, L4 Supervise and L5 Autonomous. Each level needs its controls in place first. At L3 an agent proposes and a person approves; L4 needs a sandbox, an agent identity, an audit log, sampled review and a kill switch.

What does an Agentic Studio project leave behind?

Four records you can hand over: a gate log of six signed decisions with linked evidence, an eval suite of graded test cases, a register row naming one human owner, and a kill switch tested before go-live. Together they are the evidence behind a Builder (C3) level.

Early access

Sign in. Start a project.

Agentic Studio opens at the Inevitable Summit, February 2027.

Join early access Sign in