AI agent platform architectureNine layers. Four planes.
The AI agent platform architecture behind Agentic Studio. A published design, ahead of opening in February 2027.
Updated
The stack
Every layer reports to every plane.
Each boundary between layers is a written contract.
- 01
Experience and surfaces
Mission Control on web, mobile, chat, IDE and CLI; an Approvals Inbox.
- 02
Channels and gateway
API gateway, Intent Gateway and a signed A2A endpoint.
- 03Agentic core
Orchestration
Conductor workflows that encode the ADLC; Wave Engine; Gate Service.
- 04Agentic core
Agent runtime
Digital workers from AgentSpecs, each run in a sandbox and worktree.
- 05
Tools and integration
Connector Fabric: MCP gateway, connector registry, token vault.
- 06
Memory and state
Working, session, episodic, semantic and procedural memory.
- 07
Knowledge and data
System Twin: census, code graph, ontology; Replay Bench datasets.
- 08
Models
Model Mesh: routing by task class, fallbacks, guard models.
- 09
Infrastructure
Kubernetes on AWS, GCP or Azure, all defined as code.
- Identity and accessAn identity per worker; short-lived tokens.
- Trust and governancePolicy, gates, ledger, kill switch.
- Observability and evalsA trace per step; the Eval Service.
- Cost and FinOpsBudgets per tenant and run.
Layers 3 and 4, orchestration and the agent runtime, are the agentic core.
Building blocks
The parts you'll rely on.
One line each. The detail lives in the architecture pack.
Gateway
Every request enters through one gateway and is clarified into a spec.
Runtime
One microVM per run, never reused across tenants.
Eval Service
Quality gates, a Replay Bench and an Evidence Pack per release.
Tracing
Every step traced with worker identity, model, prompt version, cost and outcome.
Policy and register
Limits live in the Policy Engine, never the prompt; an append-only, hash-chained Trust Ledger.
Tenant isolation
Tenant context comes from the verified token, never the prompt.
Data residency
A tenant is pinned to a home region; AWS Mumbai (ap-south-1) for India.
SSO
Your identity provider via OIDC, with SCIM for joiners and leavers.
Deployment
Where does your content live?
Prompts carrying your code and data never pass through the Control Plane.
Decides and records
Control Plane
Metadata, policy, gates, the ledger
Acts and holds content
Execution Plane
Sandboxes, tenant content, traces
Reached only through the Connector Fabric
Customer systems
Repos, SaaS, clouds
Four deployment views, chosen per tenant: multi-tenant SaaS, dedicated, your cloud, or on-premises. Read about India data residency and the kill switch.
Quick answers
Questions, answered.
How is Agentic Studio built?
As nine layers, from experience surfaces down to infrastructure, crossed by four planes: identity and access, trust and governance, observability and evals, and cost. Orchestration and the agent runtime form the agentic core. It is a published design today, ahead of opening in February 2027.
How are AI agents evaluated before release?
Through the Eval Service. Each agent has graded test cases that grow with every spec, a Replay Bench for regression checks and an Evidence Pack per release. G5, the quality gate, asks whether it works, safely, at load, before governance and go-live at G6.
Where do agent policy limits live?
In the Policy Engine, never in the prompt. Autonomy limits are enforced at run time, approvals go to a people-facing inbox, and every decision is written to an append-only, hash-chained Trust Ledger. That way a prompt cannot widen what an agent is allowed to do.
How are tenants isolated?
Each run gets its own microVM, never reused across tenants. Tenant context comes from the verified token, not the prompt, and each tenant is pinned to a home region: AWS Mumbai (ap-south-1) for India. Your content stays in the Execution Plane, outside the Control Plane.
Early access
Sign in. Start a project.
Agentic Studio opens at the Inevitable Summit, February 2027.