OrchStack Runtime

Execute Workflows with Total Reliability

The battle-tested execution engine behind every OrchStack agent. Full tracing, 5-tier memory, sandboxed tool execution, and human-in-the-loop gates — so your agents never lose state and always stay under control.

Orchestration Engine

Built-in Orchestration Patterns

Four battle-tested patterns for composing agent workflows. Mix and match to model any business process.

Sequential Chain

Tasks execute one after another. Output of each step feeds into the next.

A
B
C

Parallel Fan-out

Fan-out to multiple agents simultaneously, then merge results at a join point.

S
A
B
C
M

Conditional Branching

Decision node evaluates conditions and routes to the appropriate branch.

IF
YES
A
NO
B

Supervisor Delegation

A supervisor agent delegates tasks to worker agents and collects their results.

SUP
W1
W2
W3
Tool Executor

Sandboxed Tool Execution

Every tool call runs in an isolated sandbox with automatic retry, timeout handling, and circuit breakers. Your agents never hang.

100+ Pre-built Connectors

CRMs, messaging, databases, payment gateways, and cloud services ready to use out of the box.

MCP Protocol Support

Connect any MCP-compatible tool server for infinite extensibility beyond built-in connectors.

Automatic Retry with Backoff

Exponential backoff with jitter ensures transient failures don't crash your workflow.

Circuit Breakers & Timeouts

Automatic timeout handling and circuit breakers prevent cascading failures across tool calls.

tools.ts
import { defineTool } from '@orchstack/sdk'

const sendInvoice = defineTool({
  name: 'send_invoice',
  description: 'Send invoice to client',
  retry: { maxAttempts: 3, backoff: 'exponential' },
  timeout: 30_000,
  sandbox: true,
  async execute(params) {
    return stripe.invoices.create(params);
  }
});
Memory Architecture

5-Tier Memory System

From ephemeral conversation context to cross-tenant collective intelligence. Every tier purpose-built for a different access pattern.

Session Memory

Redis

Conversation context within a single interaction. Fast, ephemeral, sub-millisecond access.

Long-term Memory

PostgreSQL

Persistent facts and learned preferences that survive across sessions and restarts.

Entity Memory

Graph Store

Structured knowledge about people, companies, and relationships between entities.

Shared Memory

Team Store

Team-wide context shared across multiple agents working on the same tenant.

Collective Memory

Federated

Cross-tenant anonymized patterns and insights that improve all agents over time.

Expanding scope from session to collective

Human-in-the-Loop

Human Gates for Critical Decisions

Your agents pause when it matters. Approval, edit, and review gates ensure humans stay in control of high-stakes actions.

Agent
Gate
Human
Continue
Abort

Approval Gates

Agent pauses execution and waits for human approval before proceeding with sensitive actions.

Edit Gates

Human can review and modify the agent's proposed action before it executes.

Review Gates

Post-execution review where humans verify outcomes and provide feedback.

Timeout Escalation

Configurable timeout with automatic escalation if the human does not respond in time.

Observability

Run Replay & Full Tracing

Every decision, every tool call, every token — traced and replayable. Debug production issues in seconds, not hours.

Every agent run is fully traced end-to-end
Step-by-step replay of decisions and tool calls
Token usage, latency, and cost per step
Export traces for debugging or compliance audits
Run Trace — run_8f3k2xCOMPLETED
0msRun Started
12msMemory Loaded12ms
45msLLM Reasoning1,240 tok820ms
870msTool: search_crm340ms
1.2sLLM Decision680 tok450ms
1.7sHuman Gate: Approve4.2s
5.9sTool: send_email180ms
6.1sRun Complete1,920 total tok6.1s total
Capabilities

Everything Your Agents Need to Execute

A complete runtime engine where every component is built to work together at production scale.

Orchestration Engine

Sequential, parallel, conditional, and supervisor patterns for composing any workflow.

Tool Sandbox

Every tool call executes in an isolated sandbox with retry, timeout, and circuit breakers.

5-Tier Memory

Session, long-term, entity, shared, and collective memory layers for persistent context.

Human Gates

Approval, edit, and review gates with configurable timeouts and escalation policies.

Run Replay

Full tracing of every decision, tool call, and token. Step-by-step replay for debugging.

Auto-scaling

Horizontal scaling from zero to thousands of concurrent runs with automatic load balancing.

Frequently Asked Questions

Ready to Deploy Production Agents?

Start executing workflows with total reliability today.

Now in early access · No credit card required