Open-source agent infrastructure

Your agent logic.
Production-grade runtime.

Deploy, operate, and secure long-running AI agents without rebuilding queues, state, sandboxes, approvals, and observability from scratch.

Start building Explore the runtime
47 checkpoints saved0 steps repeated100% state recovered
RUN / RM-8F21 LIVE
RESEARCH AGENT

Competitive landscape

RUNNING
Run createdresearch-agent / prod
Model responsestep 46 committed
Tool completedbrowser.search
Checkpoint 047state persisted
Worker interruptedinstance lost
Run resumedstarting at step 48
Failure recovered automaticallyCheckpoint 047 loaded in 1.8s
BRING YOUR STACK
OpenAI Agents SDK LangGraph CrewAI Custom loops Python TypeScript
01 / THE OPERATIONS GAP

Building an agent is easy. Operating one is not.

A model, prompt, tools, and a loop are enough for a demo. They are not enough for customer workflows that run for minutes, hours, or days.

RunMesh unifies the infrastructure around your agent logic so your team can ship reliable workflows instead of maintaining an internal platform.

WITHOUT RUNMESH

Months of infrastructure work

QueueDatabaseContainersTracingApprovalsGateway
WITH RUNMESH

One operational runtime

RunMesh runtimeDurable • Secure • Observable
02 / RUNTIME PRIMITIVES

Everything between start and done.

Production controls that wrap any agent framework without taking over your agent logic.

01

Durable execution

Every meaningful action is checkpointed. Runs survive provider outages, worker crashes, deploys, and restarts.

Resume, don’t restart →
02

Isolated sandboxes

Give each run its own files, browser, tools, packages, limits, and customer-scoped resources.

One sandbox per run →
03

Complete observability

Inspect model calls, tool results, state changes, errors, retries, costs, and human decisions in one timeline.

Replay every event →
04

Human approvals

Pause sensitive actions for review without losing state. Continue from the exact approval boundary.

Policy-driven control →
05

Cost controls

Set per-run budgets, timeouts, retry limits, and model policies before an agent starts spending.

Predictable by default →
03 / DURABILITY FIRST

Step 47 failed.
Step 48 still happens.

When a worker disappears mid-run, RunMesh restores the latest durable checkpoint, rehydrates the workspace, and continues—without replaying completed model or tool calls.

  • ✓ Exact state recovery
  • ✓ Idempotent retries
  • ✓ Auditable event history
EXECUTION TRACE24H TIMEOUT
✓
STEP 46tool.result
◆
CHECKPOINT 047state persisted
×
WORKER LOSTinfrastructure failure
1.8s recovery
↻
STEP 48run resumed
04 / FIVE LINES TO DURABLE

Wrap the agent you already have.

Add the SDK, choose your operational policy, and run. RunMesh handles state, retries, approvals, costs, and recovery around your code.

1 Install the SDK2 Add the decorator3 Ship the agent
agent.py
from runmesh import agent_run

@agent_run(
    timeout="24h",
    retries=5,
    budget_usd=20,
    approval_policy="production"
)
async def research_agent(task):
    ...
READY TO DEPLOYrunmesh deploy →
OPEN SOURCE / DEPLOY ANYWHERE

Your agents. Your cloud.
Your operational control.

RunMesh is built in the open and designed for managed cloud or self-hosted deployment inside your own infrastructure.

Read the quickstart →