- Home
- Security
Autonomy is earned. Everything is on the record.
You decide how far each agent may go. You can stop all of them at once. And afterwards you can show exactly what happened, in an order nobody can quietly rewrite.
Four levels. Each allows one more class of action.
Every agent has a level from 0 to 3. Library agents default to Act; sensitive roles such as clinic, school and privacy agents ship at Assist.
| Level | Runs on its own | Waits for you | In practice |
|---|---|---|---|
| 0 · Observe | read | fetch · write · external · critical | Reads and reports. Every change needs your approval. |
| 1 · Assist | read · fetch | write · external · critical | Researches freely. Asks before changing anything. |
| 2 · Act default | read · fetch · write | external · critical | Works inside the platform. Asks before touching the outside world. |
| 3 · Autonomous | read · fetch · write · external | critical | Acts end to end. Only critical actions wait for you. |
Every tool says how dangerous it is.
Risk is attached to the tool, and sometimes to how it is called: an HTTP GET is a fetch, a POST is external. Tools that arrive over MCP are treated as external unless you say otherwise.
- read
- Internal reads: memory, files, the clock, maths, asking a coworker.
- fetch
- Outbound GET requests to the internet. A separate class because URLs can leak data.
- write
- Internal changes: memory, files, tasks, delegation, notifications.
- external
- Side effects outside the platform: POSTs to APIs, messages, MCP tools by default.
- critical
- Money, deletion and arbitrary code. No autonomy level covers it: it waits for a human unless you have set a standing rule for that exact tool.
Say yes once. Or every time.
- Durable approvalsA paused step is checkpointed. It can wait for days, through restarts, until you decide.
- Approval steps fail closedAn Approval node that times out takes the rejected path. Silence is never consent.
- Standing rulesTurn a decision into a rule: “always allow web_fetch for Riya”, or deny a tool for every agent.
- BudgetsPer-run token and step ceilings (defaults: 400,000 tokens and 200 steps) and a cap on concurrent runs.
One switch. Everything stops.
The kill switch is a hard stop: no model calls, no tool calls, and running work pauses where it is. Because every step is checkpointed, turning it back off resumes cleanly.
Separately, the always-on switch pauses only the autonomous side, the scheduler, heartbeats and cron triggers, so agents work only when asked.
Checking live status…
This chip reads the real state of this instance: always on, scheduler paused, or kill switch on.
- Kill switch
- Stops model calls and tool calls; runs pause.
- Always-on off
- No cron triggers or heartbeats; manual runs still work.
- Pause one agent
- A paused agent refuses new work until resumed.
An append-only log that can prove it hasn’t been edited.
Each entry stores the SHA-256 hash of the previous entry plus its own content. Change or delete any line and every hash after it stops matching, and verification names the first broken entry.
Illustrative entries. The first entry links to a genesis hash of 64 zeros.
Some data never leaves the room.
Runs, agents and nodes carry a data class, and the strictest one wins. Every model provider declares a region. With the guard on (the default), the router only uses a model in an allowed region, or refuses the call.
| Data class | Models allowed in |
|---|---|
| public | us · in · eu · local |
| internal | us · in · eu · local |
| confidential | us · in · eu · local |
| personal-in · Indian personal data | in · local |
| children · minors’ data | in · local |
Small rules that close big holes.
- AI disclosureAgents that speak to people outside the company say they are AI, in line with Article 50 of the EU AI Act.
- SSRF guardAgents may only reach public internet hosts over http(s). Private and internal addresses are refused.
- Code in isolationCode steps run in a separate worker with a hard timeout, and running arbitrary code is a critical action.
- Loopback by defaultA fresh install answers only on its own machine. Exposing it requires a token.
- Strict browser policyA strict Content-Security-Policy: no inline scripts, no third-party requests, no framing. This site follows it too.
- Zero runtime dependenciesNo npm packages at runtime means no supply chain to poison. Node.js and SQLite, and our code.
Designed to support the rules you answer to.
The controls above map onto obligations Indian and global businesses face. They help you meet them; they don’t make you compliant on their own, and Shiva Agents is not certified against any of these.
- DPDP Act, India
- Data classes and the residency guard for Indian personal data and children’s data; a privacy agent to keep practices current with the Act and its 2025 Rules.
- EU AI Act
- AI disclosure for agents talking to outsiders (Art. 50); human oversight through autonomy levels and approvals; records through the audit log.
- ISO/IEC 42001 · NIST AI RMF
- An AI-governance agent that maps your AI systems to these frameworks and keeps the records current.
- Incident clocks
- An incident-response workflow that triages failures and tells you which reporting clock applies, including CERT-In and DPDP.
Hire the people who keep the others honest.
The trust-and-safety blueprint brings a privacy officer, an AI-governance lead, a security analyst, an incident commander and an access auditor.