1. Home
  2. Platform
Platform

One runtime for agents that never clock out.

A single Node.js process with built-in SQLite and no npm runtime dependencies. It schedules, remembers, checkpoints, governs and routes, so your agents can work for weeks without you babysitting them.

Architecture

Everything an agent does passes one gate.

Surfaces at the top, the company and its engine in the middle, and a governance gate in front of every model call, tool call and memory write.

Shiva Agents runtime architecture Seven surfaces (web dashboard, Android app, OpenAI-compatible API, MCP, A2A, webhooks, Server-Sent Events) feed one Node.js runtime. Inside it a scheduler wakes agents and fires workflow triggers; a company of agents (virtual actors with inboxes, heartbeats and standing orders) and a durable workflow engine assign work to each other. Every action passes a governance gate before it reaches memory and state, tools, or the model router, which picks the AskShiva GPU, any OpenAI-compatible provider or an offline simulator. OpenTelemetry-style traces, Server-Sent Events and a hash-chained audit log make it observable. SURFACES Web dashboarddesign · operate Android appcompany, mobile /v1 APIOpenAI-compatible MCPserver + client A2Aagent card WebhooksIdempotency-Key SSE streamruns + tokens SHIVA AGENTS RUNTIME · ONE NODE.JS PROCESS · NO NPM RUNTIME DEPENDENCIES Scheduler cron · Asia/Kolkata agent heartbeats delayed steps retries · timeouts always-on switch wakes fires Company of agents virtual actors: asleep in the database, awake on events inbox (Karya) · heartbeats · standing orders (Sankalpa) CS CO GT WH ··· assign · delegate · consult Durable workflow engine every step checkpointed · resumable · fork from any step Team node: SOP · chat · debate · crew · role-play · handoff · ledgers Governance gate autonomy × risk · approvals · always-allow rules · kill switch · residency · SSRF guard · AI disclosure Memory and state core blocks · ranked long-term recall company knowledge · nightly consolidation SQLite: runs, checkpoints, memory, audit Tools read · fetch · write · external · critical built-ins + outside tools over MCP public hosts only (SSRF guard) Model router AskShiva GPU · Ollama · Gemma 4 · Qwen 3.6 any OpenAI-compatible provider offline simulator, clearly labelled OBSERVABLE OpenTelemetry GenAI-style traces · Server-Sent Events · hash-chained audit log · token and cost accounting
Fig. 1 · The runtime. Solid blue: the gate every action passes. Saffron: always-on signals.
Agent Studio

An agent is a persona with a job, not a prompt.

Describe the role in plain language and the platform drafts the whole agent. Every field stays editable, and every agent is the same kind of object whether you built it or hired it.

Start from a library persona

Persona
Role, goal, backstory and working style. The style is how it writes; the goal is what it optimises.
Tools
Memory, web, files, coworkers, HTTP and anything you connect over MCP. Each tool carries a risk class.
Memory
Core memory blocks it always sees, plus long-term memory it searches when it needs to.
Autonomy 0–3
Observe, Assist, Act or Autonomous: how far it may go before it asks you.
Heartbeat
A cron schedule (IST by default) on which it wakes to work its inbox.
Sankalpa
Standing orders: recurring duties with their own schedule, like “write the owner a morning briefing at 08:30”.
Verifier
An optional judge that checks the work against criteria before it moves on.
Describe → draft
Write “a GST specialist who reconciles ITC every month and flags mismatches” and get a complete first draft.
Durable execution

Runs that survive restarts, weekends and you.

Agent work is slow, expensive and sometimes waits on a human for days. The engine treats every run as something that must not be lost.

  • Checkpoint every stepEach step’s result is saved before the next one starts, so nothing is computed twice.
  • Resume after restartsDeploy, crash or reboot: waiting and running work picks up where it stopped.
  • Approvals that wait for daysAn approval or question can sit as long as it needs to; approvals fail closed on timeout.
  • Time-travel forkFork a finished or failed run from any step, change an input and replay from there.
  • Retries and timeoutsSo one flaky API or slow model doesn’t sink a whole run.
  • Loop guards and budgetsBounded loops plus per-run token and step ceilings (defaults: 400,000 tokens, 200 steps).
Team patterns

Seven ways for agents to work together.

The Team node runs topologies that research and open-source frameworks have already proven. Pick one per step; mix them across a workflow.

PatternHow it runsLineage
Assembly line SOPRoles run in order; each hands a typed artifact to the next, following written steps.MetaGPT · ChatDev
Group chatMembers share one conversation; a selector chooses the next speaker and a chair closes.AutoGen / AG2 · AgentScope
Debate + judgeMembers argue opposing cases over rounds; a judge decides and records the dissent.Shiva war room · AgentScope
Manager + crewA manager plans, delegates subtasks to specialists and synthesises the result.CrewAI hierarchical
Role-play pairAn instructor persona directs an assistant persona one step at a time.CAMEL
Handoff networkThe first agent answers or hands the whole conversation to the right specialist.OpenAI Swarm · Agents SDK
Orchestrator ledgersA task ledger and a progress ledger; the orchestrator directs specialists and re-plans when stuck.Magentic-One
Node catalogue

Twenty-two building blocks.

Familiar workflow nodes, plus the ones agents need: teams, verifiers, approvals, memory and the company inbox.

Triggers

  • Manual / formStart from the dashboard, the phone app, the API or another agent, with an optional input form.
  • ScheduleCron, always-on. Asia/Kolkata by default.
  • WebhookPOST /hooks/<id> with a secret. An Idempotency-Key fires once per key.
  • Company eventA task created or done, a run failed, an approval decided.

Agents

  • AgentOne agent works the task with its persona, memory and tools; an optional verifier checks it.
  • TeamSeveral agents in one of the seven topologies above.
  • Fan-out (map)Run an agent on every item of a list in parallel, then collect.
  • Model callA single prompt without persona or tools: extract, rewrite, classify.
  • Assign to the companyPut a task in an agent’s inbox; its heartbeat picks it up.

Logic

  • BranchRoute by rules, or let a model classify the input.
  • VerifyA judge returns pass, partial or fail, each its own route.
  • MergeWait for every live branch, then combine their outputs.

Humans

  • ApprovalStop and ask the owner: approved or rejected. Fails closed on timeout.
  • Ask a humanAsk a question and carry on with the answer.

Data

  • Set stateWrite to the run’s shared blackboard.
  • CodeTransform data with JavaScript in an isolated context.
  • MemorySearch or write company knowledge or an agent’s long-term memory.

Integration and flow

  • HTTP requestCall any public API, with templates in URL, headers and body.
  • NotifySend a notification to the dashboard and the phone app.
  • WaitPause a branch; the run survives restarts while it waits.
  • Sub-workflowRun another workflow and continue with its output.
  • ResultSet the run’s final result for the API, parent workflow and app.
Always-on company

Thousands of agents. Zero cost while they sleep.

Agents are virtual actors: an idle agent is just a row in the database. The scheduler and company events wake exactly the agents that have work, then let them sleep again.

  • Inbox (Karya)Every agent has a task inbox. Workflows, people and other agents drop work into it.
  • HeartbeatsEach agent wakes on its own cron, in IST by default, to work through what’s waiting.
  • Standing orders (Sankalpa)Recurring duties with their own schedule: the morning briefing, the weekly competitor check.
  • Delegate and consultAgents list their coworkers, hand off tasks and ask each other for an opinion.
  • Company eventsA finished task, a failed run or a decided approval can wake the next agent or workflow.
  • One switchTurn always-on off and agents only work when asked. The kill switch stops everything.
Memory

Agents that remember what worked.

Memory is modelled on Letta: small core blocks the agent always sees, and a large long-term store it searches and writes to as it works.

Core blocks
Persona, owner, company and scratchpad. Agents edit them with tools as they learn.
Long-term memory
Episodic (what happened), semantic (what is true) and procedural (how to do it).
Recall
Full-text search ranked by relevance, importance and recency.
Company knowledge
A shared store every agent can read, so one agent’s lesson helps the rest.
Consolidation
Every night, the day’s memories are folded into durable company knowledge.
Models and protocols

Bring your GPU. Speak every protocol.

The model router picks a provider per call, respects the residency guard and the kill switch, and falls back to an offline simulator when nothing is reachable. Agents are reachable over the protocols your stack already speaks.

  • OpenAI-compatible /v1: call any agent as a model, agent:<key>
  • MCP server and client; A2A agent card at /.well-known/agent-card.json
  • Webhooks with idempotency, Server-Sent Events, OpenTelemetry GenAI-style traces

Models on this instance

Checking…

  • Gemma 4 · 26Bgemma4:26b-a4b-it-qat · writing, English and Hindidefault
  • Qwen 3.6 · 35Bqwen3.6:35b · reasoning and visionreasoning
  • Any OpenAI-compatible providerconfigure a base URL, key and modelsoptional
  • Offline simulatorkeeps flows running, clearly labelledfallback

When this instance is online, the list shows its configured models and whether each is healthy.

See it run

The fastest way to understand it is to watch one run.

Open the app, run a pattern with its example input, and step through the checkpoints.