Integrating OpenSentience with Graphonomous, Delegatic, AgenTroMatic, Agentelic, FleetPrompt, and WebHost.Systems.
OpenSentience is designed to work standalone, but its full value emerges when integrated with the [&] ecosystem. Each integration point adds a governance dimension that the shim enforces at runtime.
The governance shim wraps a Graphonomous instance as a governed agent, adding graph_access permissions to control which agents can query or modify the knowledge graph.
OpenSentience.install_agent(%{
agent_id: "graphonomous-primary",
name: "Primary Knowledge Graph",
child_spec: {Graphonomous.Application, []},
permissions: %{
granted: [
%{type: :graph_access, access: :read, resource: "*"},
%{type: :graph_access, access: :write, resource: "*"},
%{type: :filesystem, access: :read, resource: "/data/knowledge.db"},
%{type: :filesystem, access: :write, resource: "/data/knowledge.db"}
]
}
})Graph access can be scoped by namespace:
graph_access:read:* — read all namespaces
graph_access:write:goals — write only to the goals namespace
graph_access:read:episodic — read only episodic nodes
This enables multi-tenant knowledge graphs where different agents have different access levels.
Delegatic is the policy authoring layer. It publishes governance policies that the PermissionEngine consumes and enforces.
Delegatic policy change
--> PubSub broadcast: {:policy_update, agent_id, changes}
--> PermissionEngine subscribes
--> ETS permission table updated atomically
--> Audit entry: policy_update from delegatic
--> Immediate enforcement on next permission checkPolicy updates are push-based via PubSub — there is no polling or TTL. When Delegatic publishes a change, the PermissionEngine receives it within milliseconds. The ETS table is the single source of truth for enforcement; Delegatic is the source of truth for authoring.
If Delegatic is unavailable, the PermissionEngine continues enforcing the last-known policy from ETS. No permissions are loosened due to connectivity issues — the shim fails closed.
AgenTroMatic implements OS-003 (Deliberation Orchestrator). The governance shim controls which agents may participate in deliberation and what tools they may invoke during the process.
%{
granted: [
%{type: :tool_invocation, access: :allowed, resource: "deliberate"},
%{type: :tool_invocation, access: :allowed, resource: "topology_analyze"}
],
denied: [
%{type: :tool_invocation, access: :denied, resource: "store_node"}
]
}This allows the agent to participate in deliberation and analyze topology, but prevents it from writing to the knowledge graph during the deliberation phase.
Agentelic publishes agent manifests — declarations of what an agent needs to run. The agent_install tool consumes these manifests directly.
Agentelic manifest
--> agent_install(manifest)
--> PermissionEngine: evaluate grants against current policy
--> Some grants may be denied if policy is more restrictive
--> AgentLifecycle: start in installed state
--> Audit: log install with manifest provenanceThe manifest declares requested permissions. The governance shim may grant fewer permissions than requested if the current Delegatic policy restricts them. The audit trail records which requested permissions were granted and which were denied, with reasons.
Agents installed from the FleetPrompt marketplace follow the same governance flow as any other agent. The marketplace manifest feeds into agent_install.
Marketplace agents start at the most restrictive defaults:
Autonomy level: observe
Permissions: only what the manifest requests, subject to policy
Lifecycle: installed (must be explicitly enabled)
The operator reviews the agent's manifest, adjusts permissions if needed, then progresses through the trust-building workflow (observe, advise, act).
WebHost.Systems is the deployment platform for the [&] ecosystem. The governance shim integrates as an OTP application within the deployed release.
The shim is configured via application environment:
# config/runtime.exs
config :open_sentience,
audit_backend: :ecto,
audit_repo: MyApp.Repo,
delegatic_pubsub: MyApp.PubSub,
default_autonomy: :observeWebHost.Systems health checks include:
PermissionEngine ETS table exists and is populated
AuditWriter GenServer is alive and responsive
AgentSupervisor is running
MCP Server is accepting connections
When multiple [&] products are governed by the same OpenSentience instance, the audit trail provides a unified view across all agents:
agent_audit(agent_id: "graphonomous-primary", type: "permission_check")
agent_audit(agent_id: "deliberation-agent", type: "action_executed")
agent_audit(agent_id: "marketplace-worker", type: "action_blocked")Each entry includes the agent_id, so cross-product queries can aggregate by agent or by event type to build a complete governance picture.