Shiva AI agents are autonomous programs designed to function as an always-on company for your business. Unlike simple chatbots, they can take a high-level goal, create a plan, use tools, and work through complex tasks until the job is done.
What are Shiva AI agents?
Most AI demos are just single loops running in a notebook. Shiva AI agents are built for the "second world"—the real world where agents must run every morning, survive server restarts, and wait days for a human manager to approve a refund.
- Autonomous Goal-Seeking: Instead of just answering a question, they take a goal, such as "Reconcile last month's GST invoices against the bank statement and flag mismatches."
- Production-Grade Reliability: They are built to be durable. This means they can handle long-running approvals and maintain a clear audit trail of everything they have done.
- Always-On Presence: Through "heartbeats" (scheduled tasks) and event triggers, agents don't just wait for you to talk to them; they can proactively work through an inbox or follow standing orders.
Reliable Agentic Architecture
To move from a simple chatbot to a reliable agent, you need more than just a large language model. You need an agentic architecture that wraps the model in a loop of ordinary, testable software.
The agent operates in a continuous loop: Goal, Plan, Act, Observe/Verify, and Remember. The model acts as the decision-maker inside this loop, but the loop itself is controlled by code that manages budgets, limits, and checkpoints.
The Core Loop Components
- Goal: The agent receives a task with a clear definition of "done," specific constraints, and a set budget.
- Plan: The agent decides the next step or drafts a full plan for the task to be reviewed.
- Act: The agent calls a tool, such as a database query, an API, or a file write.
- Observe and Verify: The agent reads the result of the tool. It asks: "Did the invoice total match?" or "Did the API return an error?"
- Remember: The agent records what happened so it can learn from its successes or failures.
Reliability is achieved through a verifier. This is a component that checks the agent's work—using code or a second model—to ensure accuracy before a task is marked complete. For example, a deterministic code check can verify if a sum of line items equals an invoice total, which is much more reliable than relying on a model's judgment alone.
Memory, Tools, and Autonomy
For an agent to work like a professional employee, it needs to remember past experiences, use company software, and know when to ask for permission.
Layered Memory
Shiva uses different layers of memory to ensure agents are efficient and informed:
- Working Memory: This is the current context window, including the immediate goal and recent steps. To keep things fast, old steps are summarized rather than kept in full.
- Episodic Memory: This is a record of past runs. It tracks what the agent tried, what worked, and what a human had to correct, preventing the agent from repeating the same mistakes.
- Long-term Memory: This holds permanent facts and documents, such as your company's price lists or customer history, which are retrieved on demand.
Tools and Interaction
Agents interact with your business via tools. These are functions the model can call, such as send_whatsapp(template_id, phone) or search_invoices(gstin, month). A good tool has a narrow scope and provides clear error messages if something goes wrong.
Levels of Autonomy
You decide how much control an agent has through risk-based guardrails. Autonomy levels range from Observe (watching and reporting) to Autonomous (acting on its own). Critical actions are gated by human approval or strict spending limits.
Visual and Durable Workflows
Shiva provides a visual canvas, similar to n8n, where you can design complex processes. In these workflows, any node can be a single agent or an entire team of agents working together.
The platform includes 20 different node types to give you total control over the process, including:
- Triggers: Start a run via a schedule, a webhook, or a manual command.
- Decision Nodes: Use branches to split workflows based on rules or model logic.
- Human-in-the-loop: Use "ask a human" or "approval" nodes to pause a workflow for manual review before a risky action is taken.
- Memory and State: Nodes that allow agents to set, get, or save information for later.
- Sub-workflows: Trigger entire other workflows as part of a larger process.
One of the most powerful features for developers is "time travel." If an agent makes a mistake or you want to test a different approach, you can fork a run from any earlier step to try a different path without starting from scratch.
Multi-Agent Team Patterns
Complex business problems often require more than one specialist. Shiva supports proven topologies that allow multiple agents to work together effectively. You can find these in the Shiva Agent Library.
| Pattern | How it works | Real-world use case |
|---|---|---|
| Assembly-line | Each role hands a typed artifact to the next specialist. | A content pipeline: Researcher → Writer → Editor. |
| Group Chat | A speaker selector picks who talks next in a conversation. | A brainstorming session between a creative and a strategist. |
| Manager-Worker | A manager agent plans tasks and assigns them to specialists. | A project manager agent coordinating a marketing campaign. |
| Handoff Network | Agents pass control to the right specialist as the task evolves. | A customer support agent handing off a technical issue to a dev agent. |
The platform provides 23 runnable workflow patterns and 9 "department-in-a-box" blueprints to help you get started immediately.
Open Protocols and Integration
Shiva is designed to be an open part of your tech stack, not a walled garden. We use standard protocols so your agents can talk to other software and other agents.
- OpenAI-Compatible API: You can call your agents using the standard
/v1/chat/completionsendpoint. Any existing OpenAI client can talk to your agents as if they were standard LLMs. - MCP Server Support: Your agents and workflows are available as tools to any Model Context Protocol (MCP) client.
- Agent2Agent (A2A) Protocols: Through A2A cards and JSON-RPC, other agents can discover your agents and delegate tasks to them automatically.
Governance and Security
Running autonomous agents requires strict oversight. Shiva provides several layers of protection to ensure agents stay within their bounds.
Security and Compliance
Data privacy is a priority. We include a data-residency guard that keeps sensitive Indian personal data and children's data away from models hosted outside of India or off your local machine.
Audit and Control
To prevent "silent edits" or hidden errors, every single action taken by an agent is recorded in a hash-chained audit log. This creates an unchangeable history of exactly what the agent did and why.
Frequently Asked Questions
- How are Shiva AI agents different from chatbots?
- Chatbots respond to messages one at a time. Agents take a goal, plan multiple steps, use tools, and work autonomously until the job is done.
- Can I approve an agent's action before it happens?
- Yes. The platform includes human-in-the-loop approval nodes, allowing you to pause risky or critical actions for manual review.
- What happens if the server restarts during an agent run?
- Shiva Agents uses durable checkpoints after every step, allowing an agent to resume its task exactly where it left off.
- How do agents remember past interactions?
- Agents use episodic memory to record what worked and what failed, preventing them from repeating the same mistakes in future runs.
- Can I use my own tools with Shiva AI agents?
- Yes. You can provide tools through function calling, allowing the model to trigger your specific code, APIs, or database queries.

Questions people ask
How are Shiva AI agents different from chatbots?
Chatbots respond to messages one at a time, whereas agents take a goal, plan multiple steps, use tools, and work autonomously until the job is done.
Can I approve an agent's action before it happens?
Yes. The platform includes human-in-the-loop approval nodes, allowing you to pause risky or critical actions for manual review.
What happens if the server restarts during an agent run?
Shiva Agents uses durable checkpoints after every step, allowing an agent to resume its task exactly where it left off.
How do agents remember past interactions?
Agents use episodic memory to record what worked and what failed, preventing them from repeating the same mistakes in future runs.
Can I use my own tools with Shiva AI agents?
Yes. You can provide tools through function calling, allowing the model to trigger your specific code, APIs, or database queries.
