Operating Boundaries
The execution, durability, evidence, action, and credential rules visible to operators and contributors.
Routine Work and Human Checkpoints
Routine work may continue without step-by-step supervision. Execution stops when it needs missing evidence, ambiguous judgment, an important decision, policy-required approval, or help resolving an uncertain external outcome.
Conversation and Work
A conversation can remain exploratory. Oblive creates a durable task when an outcome needs explicit ownership, scheduling, dependencies, retries, artifacts, or follow-through.
Durable State
Important state must survive a lost process, dropped notification, duplicate wakeup, or restart. The browser, a model session, Redis transport, and a temporary workspace are never the only owners of operational state.
Evidence
Organization context retains sources and citations. Agents distinguish supported facts, inferences, assumptions, and missing information instead of presenting all four as equally certain.
Consequential Actions
External changes are typed actions with an idempotency contract, policy state, and receipt. When an outcome cannot be confirmed safely, the action moves to reconciliation instead of being retried blindly.
Credentials
Provider secrets are encrypted and resolved by the backend only for the scoped execution that needs them. They do not belong in prompts, manifests, workspaces, logs, or general agent environments.
Extension Points
New capabilities use the existing catalog, skill, MCP, CLI, action, API, and profile boundaries. Avoid adding duplicate sources of truth or services without a durable ownership requirement.
Contributors should read the Architecture Guide before changing a lifecycle, durable owner, external effect, or runtime boundary.