Agents get job descriptions.
One Use Case, one persona, tested Stories, scoped authority, a task branch, and a token budget. Every result is reviewable evidence.
The session is the security context.
Bindings resolve at admission and stay immutable. A different role is a different session.
Token budget: the maximum Legra-token spend authorized for this task across priced data, compute, storage, transport, rights, and services. See how token pricing works.
Modeled business intent becomes the agent's capability surface.
Legra derives the agent's world from governed graph objects that people inspect, test, version, and improve.
Four levels make the trade-off explicit.
Every deployment names its level. The server enforces the boundary; prompts, branch defaults, and confirmation dialogs decorate it.
Preferred
Persona-bound Use Case harness
A dedicated Legra harness receives the Stories for one persona in one root Use Case and its descendant tree.
Use Case, persona, deployment, authority, token budget, task, and branch are immutable session bindings.
Strong
Sandboxed standard harness
A standard agent harness runs in a VM or container and receives the same server-enforced, Story-only capability as level 1.
Semantic scope remains in Legra; the sandbox adds process, filesystem, resource, and network isolation.
Branch-safe
Full freedom on one staging branch
A sandboxed harness gets full SPARQL, Cypher, or GQL access. The server pins it to one staging branch forked from the default branch.
The credential stays on its branch. A human or a separate approval agent decides promotion.
High risk
Broad workspace statement access
The agent issues raw statements across every branch its workspace credential reaches.
The Journal, provenance, and graph history show what happened and restore the original state. The authority is the risk.
Autonomy runs inside a reviewable path to production.
01
Bind
Admit one principal, deployment, root Use Case, persona, task, token budget, and branch.
02
Derive
Build the exact Story tool surface, schemas, concepts, instructions, and skills from the graph.
03
Work
The agent invokes approved capabilities and writes proposed knowledge to its task branch.
04
Prove
Attach calls, costs, inputs, outputs, changes, and external activity to one Task and Journal evidence tree.
05
Promote
Validate Stories, shapes, contracts, and budgets; review the graph diff; merge only approved work.
Independent boundaries compose.
Semantic scope bounds intent. Authority bounds capability. Branches bound effect. A failure in one layer stays inside the others.
Semantic scope
One root Use Case resolves to one governed descendant tree. The caller gets that tree and nothing beside it.
Authority
The agent receives attenuated, short-lived capability. Workspace membership stays with people.
Effect containment
A task branch isolates proposed writes. Read grants and Story scope bound reads.
Execution and egress
A sandbox limits processes and resources. Deployment data sources name the reachable destinations.
Evidence and economics
Tasks and the Journal connect every action to its actor, cost, approval, and recoverable graph state.
The harness turns governed graph context into executable authority.
The hosted harness generates the persona package, resolves the Use Case tree and Story surface, binds the task branch and token budget, and runs the agent inside that scope. The CLI MCP server is the supervised operator surface.