Niyatna OS & Workspaces
How Niyatna OS isolates agent execution, manages credentials, and stages verification proofs for the human Command Room.
An Agentic Company runs on specialized agent workforces. To ensure complete security, process integrity, and operational alignment, every Niyatna Agent operates in a dedicated Workspace managed by Niyatna OS.
The architecture is built around one core principle: the human supervisor retains total visibility and veto power, while the agent runtime executes in secure isolation.
Sandbox Runtime Mechanics
Each specialized agent (e.g., CFO or COO Agent) runs inside a jailed runtime container. Niyatna OS controls the boundaries:
- Strict Isolation: Agents can only read and write to directories explicitly mounted to their workspace container.
- Ephemeral Life: Workspaces are spawned dynamically when an intent packet is queued, and archived once verification is complete.
- Real-Time Auditing: Every terminal command, code execution, filesystem change, and API request is logged to an immutable proof trail.
Niyatna HQ Interface
Human operators inspect and direct agent work through Niyatna HQ. The interface is divided into four key panels:
- The Intent Rail: Pinned at the side, displaying the active goal, constraints, and success criteria.
- The Logs Feed: A live audit stream showing the agent's reasoning, tool calls, and execution steps.
- The Proof Panel: Displays the outcomes—such as database schema diffs, CRM sync logs, or payment audit receipts.
- The Decision Gate: Staging area showing high-risk actions requiring explicit human approval before being executed.
Security Architecture
Traditional scaling requires sharing sensitive API keys and database login credentials across multiple teams. Niyatna OS solves this by separating credentials from execution:
- Scoped Token Injection: LocalRoute manages master credentials and issues temporary, scoped session tokens to the agent sandboxes.
- Zero Write Persistence: Sandbox containers do not persist changes to the server unless they pass the Verification Layer and receive human approval.
- Veto Boundaries: Irreversible actions (like running database migrations or triggering customer-facing notifications) are blocked by default and require operator validation.