We started Everruns because we kept watching the same thing happen: someone would build an impressive AI agent — a research assistant, a code reviewer, an automated analyst — and it would work beautifully for five minutes. Then it would crash. And everything it had done was gone.

This isn't a niche problem. It's the central bottleneck standing between the current generation of AI demos and the next generation of AI that actually works.

## The Gap

AI capabilities are advancing fast. Context windows are growing from thousands to millions of tokens. Reasoning is getting deeper. Tool use is expanding. Agents can now browse the web, write code, run experiments, and coordinate multi-step workflows autonomously.

But the infrastructure running these agents hasn't kept up. It's still the same story: ephemeral processes on machines that crash, networks that drop, and containers that get recycled. When your agent finishes in 30 seconds, that's fine. When it's been working for 8 hours and the machine restarts, you've lost everything.

Research from [METR](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/) shows AI task completion capabilities are doubling every seven months. The industry is building agents that can work for hours and days. But nobody is building the infrastructure to actually run them reliably.

That's the gap we exist to close.

## What Everruns Is

Everruns is a durable execution platform for AI agents. Every tool call, every LLM response, every state transition is automatically persisted. If the machine crashes after step 47 of a 100-step workflow, the agent resumes at step 48 — not step 1.

This isn't new computer science. Databases solved this with write-ahead logs decades ago. Distributed systems have consensus protocols. Message queues have delivery guarantees. We're bringing the same reliability guarantees to AI agent execution.

What makes agents different from traditional workloads:

- **Non-deterministic outputs.** You can't replay an LLM call and expect the same answer. Durable execution for agents must persist results, not just replay operations.
- **Large, evolving state.** Agent context includes conversation history, tool results, accumulated knowledge — megabytes of data that grows over time.
- **Real-world side effects.** Agents send emails, write files, call APIs. These can't be undone on retry.
- **Unpredictable timing.** One tool call takes 50ms, the next takes 5 minutes. The system can't assume anything about execution duration.

We built Everruns to handle all of this natively.

## Why Open Source

We believe the agent execution layer is infrastructure, not a product feature. Just as Linux, Kubernetes, and PostgreSQL are open infrastructure that everyone builds on, the agent runtime should be too.

Open source means:

- **You own the platform.** Run it on your machines, in your security perimeter, under your governance policies.
- **No vendor lock-in.** Your agents aren't trapped in someone else's cloud.
- **Community-driven.** The people building agents are the best people to shape the infrastructure those agents run on.

The entire platform is [on GitHub](https://github.com/everruns/everruns). MIT licensed.

## Where We're Headed

We see three horizons for durable agent execution:

**Now: research and analysis agents.** Agents that spend hours or days gathering information, processing data, and synthesizing findings. These are the first workloads where durability changes everything — because the cost of restarting from scratch is measured in hours of lost work and wasted API calls.

**Next: enterprise headless agents.** Backend workflow automation — processing orders, coordinating systems, managing data pipelines. These agents need the same reliability guarantees as the databases and message queues they interact with. No silent failures. Full audit trails.

**Beyond: physical-world agents.** Laboratories, factories, logistics networks. AI agents directing experiments, monitoring equipment, controlling robotic systems. Here, a lost agent doesn't just waste compute — it wastes physical materials and lab time. Durable execution becomes safety-critical.

## What's Next

For the full picture of what we believe and where this is going, read our [Vision](/vision/).

We're shipping in the open. Follow along on [GitHub](https://github.com/everruns/everruns), or reach out at [contact@everruns.com](mailto:contact@everruns.com) if you're building agents that need to run longer than a few minutes.

We'll be writing here about what we're learning — the hard problems of agent infrastructure, the design decisions we're making, and the patterns we see emerging as AI agents get longer-running and more autonomous.

This is the beginning.