Polos – Open-Source Runtime for AI Agents

TL;DR

TL;DR: Polos is an open-source runtime that handles the hard production problems for AI agents — sandboxing, durable execution, approvals, and observability — so you only write the agent logic.

What Is Polos?

Polos is an open-source runtime for AI agents. It handles the infrastructure problems that make shipping agents to production hard: isolated sandboxed execution, durable workflows that survive crashes, human-in-the-loop approvals, and built-in triggers.

The project was shared on Show HN in mid-2026 by ndeodhar.

Architecture

Polos has two main components:

• Orchestrator – Written in Rust. Manages execution state, durable logs, retries, scheduling, triggers, and the dashboard UI. Backed by Postgres.
• Worker – Runs your agents and workflows, written with the Python or TypeScript SDK. Connects to the orchestrator. Scale horizontally by running multiple workers.

Setup Workflow

Step 1: Install the CLI

curl -fsSL https://install.polos.dev/install.sh | bash

Step 2: Create a New Agent Project

npx create-polos          # TypeScript
pipx run create-polos    # Python

The CLI prompts you for a project name and LLM provider (Anthropic, OpenAI, Google, etc.).

Step 3: Run the Agent

cd my-project && polos dev
polos run coding_agent

The dashboard is available at http://localhost:5173.

Deeper Analysis

Sandboxed Execution

Agents run in isolated Docker containers (or E2B, cloud VMs). Every sandbox ships with built-in tools — exec, read, write, edit, glob, grep, web_search — so your agent can run commands, navigate codebases, and browse the web out of the box. No tool code to write, no sandbox lifecycle to manage.

Durable Workflows

If your agent crashes mid-execution, Polos replays the workflow from the durable log, returning previously-recorded results instead of re-executing them. Completed steps are never re-executed — you never pay for an LLM call twice. Prompt caching delivers 60–80 percent cost savings on retries.

Human-in-the-Loop

Any tool call can be routed through an approval flow. Reach your team via Slack, Discord, or email. Configurable rules control what needs approval. Paused agents consume zero compute.

Triggers

Every agent gets a webhook URL out of the box. Point GitHub, JIRA, Salesforce, or any system at it. Plus HTTP API, cron scheduling, event-driven triggers, and built-in Slack integration to run agents from chat.

Observability

OpenTelemetry tracing for every step, tool call, and approval. Full execution history. Visual dashboard for monitoring and debugging.

Bring Your Stack

Build agents on Polos with any LLM — OpenAI, Anthropic, Google, and more — via Vercel AI SDK and LiteLLM. Or bring existing agents from CrewAI, LangGraph, and Mastra and get Polos’s durability, observability, and sandboxing without rewriting anything.

Practical Evaluation Checklist

• [ ] Run the install script and verify CLI is in PATH
• [ ] Create a project with npx create-polos or pipx run create-polos
• [ ] Start the server with polos dev and confirm localhost:5173 loads
• [ ] List available agents with polos agent list
• [ ] Run polos run <agent> and observe sandboxed execution
• [ ] Inspect the dashboard for tracing output
• [ ] Test approval flow configuration

Security Notes

• Sandboxed Docker execution prevents agents from accessing host resources by default
• Approval workflows gate dangerous operations (exec, write to sensitive paths) before execution
• Durable logs are stored in Postgres — secure the database access accordingly
• Review sandbox tool permissions before giving agents broad filesystem access

FAQ

Q: Who built Polos? A: The project is maintained by polos-dev on GitHub. It was launched on Show HN by ndeodhar.

Q: What license does Polos use? A: Apache 2.0, confirmed via the LICENSE file in the repository.

Q: Can I use existing LangGraph or CrewAI agents with Polos? A: Yes. Polos supports bringing agents from CrewAI, LangGraph, and Mastra without rewriting them, gaining Polos’s durability, sandboxing, and observability.

Q: What LLMs does Polos support? A: Any LLM supported by Vercel AI SDK or LiteLLM — OpenAI, Anthropic, Google Gemini, and more.

Q: How does durability work under the hood? A: Polos captures the result of every side effect (tool calls, API responses, time delays) as a durable log. On crash, it replays the workflow from the log, restoring exact local variables and call stack.

Conclusion

Polos solves the gap between “agent works in a notebook” and “agent runs reliably in production.” By handling sandboxing, durable execution, approval flows, triggers, and observability as built-in primitives, it lets you focus on writing the agent logic rather than the infrastructure around it. If you are building agents that need to survive production workloads — handle failures gracefully, run in isolation, and give humans a way to approve sensitive actions — Polos is worth evaluating.

Source and Accuracy Notes

• Project page: polos.dev
• Source repository: github.com/polos-dev/polos
• License: Apache 2.0 (verified via LICENSE file)
• HN launch thread: news.ycombinator.com/item?id=47153680
• Source last checked: 2026-07-07 (commit verified against main branch)