entergenie.Let’s talk
DocsCore model

Core model

The building blocks of work in Entergenie, from agents and tasks to policies and guardrails.

A shared language for real work

These concepts connect intent, execution, access, and quality. Use the same model through the UI, CLI, and Terraform, alongside chat, APIs, and developer tools.

ConceptDefinitionExample
AgentAn AI worker configured to perform a specific type of work.Incident Investigator
TaskA goal assigned to an agent, including its expected outcome and constraints.Investigate elevated API errors
SessionA single execution of an agent working on a task. A task may have multiple sessions.Investigation session #123
ToolAn action an agent can invoke, potentially affecting external systems.search_logs
ConnectionA configured integration providing access to an external system, typically used by tools.Production logs MCP
HarnessThe combined runtime and supporting mechanisms that enable an Agent to perform work.Agent runtime with sandbox
GuardrailsGuidance and checks that steer and verify an Agent’s work against defined norms.Architecture rules, approved fixtures, structural tests, linting
DelegationAssignment of a task from one agent to another. Does not automatically grant authority.Investigator delegates log analysis
PolicyRules governing permitted actions, limits, and required approvals.Read-only production access

How the pieces fit

How the Entergenie core model fits togetherA task is assigned to an agent, which executes it in a session. A principal issues a delegation grant that limits the session’s execution identity. Inside the harness, policy checks both the session’s requested action and its effective authority before tools can run. Tools use connections to access external systems.WORK & EXECUTIONIDENTITY & AUTHORITYHarnessIsolated runtime · policy enforcementassigned toissuesexecutes inbounds accessactionauthorityuseprovide access toTaskGoal, outcome, and constraintsPrincipalPerson or organizational identityAgentConfigured AI workerDelegation GrantScoped, limited authorizationSessionOne execution of the taskExecution IdentityEffective identity and authorityPolicy checkPermitted actions, limits, approvalsToolsActions an agent can invokeConnectionsConfigured, authenticated accessExternal systemsYour business systems and services
Work and tool access Delegated authorityScroll to explore on smaller screens.

A Task describes what needs to happen. An Agent provides the configured worker. A Session is one execution attempt inside the Harness. Tool calls are checked against Policy and the session’s Execution Identity before a Connection supplies access to an external system.

Guardrails and policy serve different purposes

Guardrails guide and check the quality of work: architecture rules, linting, or an evidence requirement. Policy controls authority: which resource can be accessed, which actions are permitted, and when approval is needed. A helpful prompt does not replace an enforced permission boundary.

Delegation is not permission

An investigator can ask another agent to analyze logs. That assignment alone must not expand access. The receiving agent needs its own valid execution authority, bounded by the relevant Principal’s grant and policy.

THE ENTERGENIE NOTEBOOK

A little signal. A lot less noise.

Ideas on agents, responsible autonomy, and getting real work done.