Build agents that fit your business.

The execution editor exposes Runtime, Model, Execution, Constraints, Memory, Instructions, Context, Tools, Routing and Trust. Start with the job and required controls; choose additional settings when the task needs them.

The instructions editor separates expertise, procedures, references and constraints. Reusable Context is assigned separately. Knowledge eligibility is not proof of retrieval, and a saved model or routing choice is not proof that the provider is configured or a run succeeded.

Use the product’s configuration screens, or integrate through the supported APIs and SDK paths. The deployment must supply the required model access, integrations and permissions. A saved draft or registered gateway does not establish a working deployment.

Inspect APIs and SDK integration → · Inspect the connected test and instruction change →

A role can publish to its own unit or an ancestor. Consumption through that placement is limited to the selected unit; publication does not automatically spread access through the organisation or grant execution authority. Existing assets without a unit target use a separate organisation-level path, still subject to access checks.

Assigning Context saves a reference to the instructions, rather than making a private copy. In the handover demonstration, Daniel authors the specialist and Avery selects the same saved Agent. Both have organisation-wide build administration, and role/trust preparation was explicit. It does not establish that every employee can discover or run the specialist.

Duplicating a pipeline creates another graph. It does not copy a complete Harness or grant additional permissions. Configuration history restores its supported saved fields; it does not undo completed work or restore an entire pipeline graph.

Shared setup and access to business information are separate. Check the intended roles, data access and execution permissions in your evaluation.

The two completed runs use the same specialist and assigned four-heading format. They demonstrate sequential execution, not parallel branches, retries or updates to operational records.

Six generic starter configurations are defined: Basic Q&A Assistant, Conversational Tutor, Research Specialist, Code Developer, Code Reviewer and Data Analyst. They still need model, permission and task-specific setup. The Northstar handover is a demonstration configuration, not a ready-to-run business application included in every installation.

Inspect the actual outputs and saved document →

1. Open Tools → MCP Servers . Enter the name, endpoint, transport and enabled state.
2. Configure the supported credential method. A secret URL can be referenced as env:YOUR_MCP_URL ; the environment variable holds the complete connection URL.
3. Save and inspect discovery. HTTP/SSE discovery runs initialization and requests the tool list.
4. Attach the server or a discovered tool to an agent, using its agent identifier and the tool name where applicable.
5. Run a task that needs the tool and inspect the actual result and permissions. Registration alone proves no business operation.

Compatibility to check: registry discovery supports HTTP/SSE. Its discovery client does not forward Authorization headers, so bearer-token or OAuth support must be verified on the chosen execution path. Stdio needs a supported runtime and operator allowlists; a stored WebSocket entry is not proof of execution support. Private-network endpoints need deployment review.

The walkthrough shows one completed individual-tool call through an external HTTP server.

The individual-tool invocation path checks the tool binding and requires approval. The Claude adapter registers the whole server, so tool exposure and approval behavior must be checked for that runtime.

The enterprise server exposes ten tools. These names describe tools, not ten separately installed servers.

Plan and route

decompose_objective , route_task

Delegate and escalate

delegate_to_agent , escalate , invoke_task_agent

Knowledge and connected systems

query_knowledge , check_directive , execute_connector_query

Return the work

create_deliverable , submit_upstream

Connector queries can also write through supported API connections; review the agent’s assigned source and permitted operation.

Inbound MCP admission is closed by default. An operator must explicitly enable it and configure the allowed origins. Clients without an Origin header need an explicit admission choice.

Transport admission is separate from authorisation. Verify the caller’s identity, organisation and permission for each operation you intend to expose. Test the actual external client and deployed configuration before allowing business work.

This describes the configuration path. The public external-tool recording shows Aegis calling an outside server; it is not evidence of a client calling into Aegis.

Knowledge publication and procedure proposals have separate review paths. Tenant learning exclusions restrict what can enter approved knowledge; a new exclusion does not automatically erase all previously published material. Procedure selection follows department/team placement, and sharing does not confer authority.

The recording above shows a proposal, edited acceptance and later use. The Controls walkthrough separately shows knowledge publication, permitted and refused access, and acknowledgement of a leadership message.

Explore access and directive controls