Propose
The agent identifies an intended action, resource and connection.
PLATFORM ARCHITECTURE
Intent. Authority.
Effect. Evidence.
A shared governance model around qualified agent interactions. The important distinction is not just what connects, but where a decision can actually be enforced.
Keep planner and runtime choices open while making the actor, authorized operation, effect and outcome explicit at the integration boundary.
Coding agentsBusiness agentsCloud workloads
One governance modelIDENTITY · POLICY · EVIDENCE
Tools & APIsKnowledge & contextInfrastructure
Shared governance. Independent identities. Explicit boundaries.
Evaluate
Current identity, policy, delegated scope and approval determine eligibility.
Act or wait
A qualified adapter performs, denies or pauses the protected effect.
Record & reconcile
Preserve evidence and the original effect identity. Resolve unknown outcomes before retrying.
BOUNDARIES ARE PART OF THE PRODUCT
Be precise about
what is controlled.
Discovery ≠ authorization. A visible tool still needs current permission when it is invoked.
Observation ≠ enforcement. A trace does not mean a path can block a request.
Configuration ≠ runtime proof. An installed artifact may not be loaded by the client.
Revocation ≠ reversal. Stopping future use does not undo a completed external effect.
Unknown ≠ no effect. A lost response does not make another attempt safe.
AN ILLUSTRATIVE GOVERNED ACTION
REQUEST → DECISION → OUTCOMEA routine action.
A routine action.
An accountable result.
An agent requests a preview deployment. Evaluate its current authority, execute within scope and retain the result.
- 01Request receiveddeploy.preview · staging
- 02Current authority checkedidentity + policy + resource scope
- 03Effect evaluatedqualified execution boundary
- 04Receipt retainedan attributable, durable outcome
01
{02
"execution": "ex_example",03
"action": "deploy.preview",04
"scope": "staging",05
"authority": "not_checked",06
"decision": "pending",07
"effect": "not_started",08
"retry_allowed": false09
}How an approved environment reaches a target.
The broader tenancy model is Organization → Workspace → Application → Environment. Within a workspace, a project selects an effective environment for its work. A profile is a versioned configuration choice and a target binding connects that choice to a supported client, account or resource; neither is an additional tenant or a grant of authority.
Copying a profile does not copy credentials or authorize a connection. Each connection needs its own account authorization and resource scope, and the actual target state must be observed before calling the environment ready.
- Workspace
- The ownership and collaboration boundary within an organization.
- Project
- A work context in the workspace that resolves an effective environment.
- Profile
- A versioned configuration selection for a supported client or workflow; it is not an authority boundary.
- Target binding
- An explicit association to the actual client, account, resource or execution target where the profile is consumed; authorization remains separate.
An entry point is not always the effect boundary.
MCP tool discovery or a qualified ACP session can help an agent reach the governance model. Neither alone proves that DriftGate can block an external write. Enforcement depends on the adapter or API path that owns the consequential operation before it happens.
At that path, carry the exact actor, resource, operation, policy revision and material approval inputs. Retain an effect identity so an uncertain provider response can be reconciled without issuing a duplicate request. Design aims to keep the pre-effect decision low latency, but this website makes no measured latency claim.
BUILD WITH INTENTION