The managed runtime for production agents.

Durable, observable, governed execution with cold starts in milliseconds and runs that last months. Use Constal SDK or bring your own framework.

TRY FOR FREE BOOK A DEMO

No. 01

COMPATIBILITY

Build in any framework. Run it on Constal.

Frameworks are how you write agents. Constal is where they run. Deploy the code you already have: no adapters, no forked SDKs, no lock-in. Every step is checkpointed at the runtime boundary, and the code that runs on your laptop is the code that runs in prod.

agent.ts — your LangGraph code, as-is

import { StateGraph } from "@langchain/langgraph";

const graph = new StateGraph(ResearchState)
  .addNode("plan", plan)
  .addNode("act", act)
  .addEdge("plan", "act")
  .compile();

export default graph;

No. 02

COLD STARTS

Milliseconds, not minutes.

Snapshot-resumed microVMs boot your agent below human perception. Scale to zero without paying for it on the way back up.

RUNTIME CLASSP50 TIME-TO-FIRST-INSTRUCTION

CONSTAL9 ms

MICROVM SANDBOX125 ms

HARDENED CONTAINER480 ms

SERVERLESS PLATFORM800 ms

CONTAINER COLD BOOT2,400 ms

Illustrative internal benchmark, August 2026.

No. 03

LONG HORIZON

Agents that outlive the process.

Every step is checkpointed to the state fabric. Agents sleep for free, wake in 11 ms, and survive crashes, deploys, and two-week human approvals.

Sleep is free.

Suspended agents cost nothing. State is pinned until a timer, webhook, or human wakes them.

Crash = retry, not restart.

Execution replays from the last checkpoint, side-effects deduplicated. No lost work, no double sends.

Human-speed gates.

Approvals can take seconds or weeks. The run waits durably and resumes exactly where it stopped.

No. 04

OBSERVE & GOVERN

Glass-box execution.

Every model call, tool call, and state change is traced by default — and every one of them passes through policy before it executes.

TRACE — run_8f3a · researcher · 31d 4h LIVE

llm.plan 320ms

tool:web.search 1.2s

checkpoint 3ms

llm.reason 640ms

tool:browser.run 2.1s

human.gate —

sleep 16d

resume 11ms

llm.synthesize 480ms

tool:report.write 1.4s

complete day 31

RBACRoles scoped down to the individual tool call

AUDITHash-chained log of every action, exportable

BUDGETSToken and dollar ceilings per agent, team, or run

POLICYAllow, deny, or gate any call behind approval

REDACTIONPII scrubbed at the runtime boundary

SSO/SCIMEnterprise identity from day one. SOC 2 in progress

No. 05

COMPOSABLE

One runtime, many agents.

Agents coordinate through the runtime, not glue code. A LangGraph planner hands off to a Mastra support agent or an OpenAI SDK researcher with typed handoffs, shared durable state, and one trace stitched across the whole architecture.

No. 06

MODULAR

Every layer is a module.

The core is opinionated: state, scheduling, and policy are the runtime. Around it, every integration point is swappable without touching the rest.

i. STATE FABRIC Checkpointed agent state, replicated across zones. Query it like a database. CORE

ii. SCHEDULER Timers, crons, webhooks, and human gates that wake agents in milliseconds. CORE

iii. SANDBOX POOL Snapshot-warmed microVMs for tool and code execution. Hard isolation per run. CORE

iv. MODEL GATEWAY One interface to every provider, with fallbacks, caching, and cost metering. OPENAI-COMPAT

v. MEMORY STORE Session, entity, and long-term memory with retention policies attached. S3-COMPAT

vi. POLICY ENGINE Declarative rules over every call: allow, deny, budget, or route to a human. CORE

vii. BYOC The full data plane, deployed into your own cloud account. COMING SOON

Your agents, running today.

TRY FOR FREE BOOK A DEMO