I found the Agent That Runs for Hours

75k★, MIT, DeerFlow: ByteDance

For weeks your feed has been one sentence on repeat: stop prompting, build the system, agents that run for hours while you sleep. Most of it is talk. DeerFlow is the part you can actually clone

It's a ByteDance project, MIT licensed, and it just crossed 75,000 stars. Version 2.0 is a ground-up rewrite that hit #1 on GitHub trending. And the one word that explains why it matters is the word the authors chose for it: harness

What a "harness" actually means

Most agent projects are frameworks - you wire the pieces together yourself. DeerFlow ships the runtime. Out of the box it gives an agent a filesystem, long-term memory, a sandbox, a library of skills, and the ability to plan and spawn sub-agents. It's built on LangGraph and LangChain, but you don't assemble it. You run make setup, answer a 2-minute wizard, and you have an agent with infrastructure

That distinction is the whole story. A chatbot with tools can call an API. An agent with a harness has a computer

The parts that matter

Sub-agents. The lead agent doesn't try to do everything in one context. It decomposes the task and spawns sub-agents on the fly, each with its own scoped context, its own tools, its own stop condition. They run in parallel, report structured results, and the lead synthesizes them. This is how one request turns into "a dozen agents researching different angles, then converging into a single report - or a website, or a slide deck." One harness, many hands. That is literally what "runs for minutes to hours" means under the hood

Its own computer. Every task gets a real execution environment - a filesystem with uploads, a workspace, and an outputs folder. The agent reads, writes, and edits files, views images, and runs shell commands inside isolated containers. Not "tool access." An actual machine it operates

Context engineering. The thing that kills long agent runs is the context window. DeerFlow fights it on purpose: sub-agents are isolated so they don't drown in each other's noise, completed sub-tasks get summarized, intermediate results get offloaded to the filesystem, and skills load progressively - only when the task needs them. That's why it stays sharp on token-sensitive models instead of melting down at message 40

Memory. Most agents forget everything the second the chat ends. DeerFlow keeps a persistent memory of your profile, preferences, and stack across sessions, stored locally. The more you use it, the more it knows your style

Run it from your phone. This is the underrated one. DeerFlow can take tasks from Telegram, Slack, Discord, Feishu, WeChat, and DingTalk - no public IP required. You text it a task from your phone, it fans out a dozen sub-agents on your machine, and texts you back the result. That's the "ship dozens of PRs from my phone" workflow people keep posting about, as an open-source box you own

The honest part

This is not a toy you leave running on a public server. DeerFlow executes system commands and reads and writes files by design, so it ships locked to localhost and the docs are blunt about it: deploy it in a trusted environment or put it behind a real auth gateway. It also wants real hardware - 8 vCPU and 16 GB RAM for serious multi-agent runs, not a 2 GB droplet. This is infrastructure, and it asks to be treated like infrastructure

Why it's worth your time

Everyone is arguing about which model is smartest. DeerFlow is a bet that the model was never the bottleneck - the harness around it was. Memory, a filesystem, sandboxed execution, sub-agent orchestration, aggressive context management. The boring infrastructure that turns a clever autocomplete into something that actually finishes a multi-hour job

If you want to understand where "agents that run for hours" actually comes from, you don't need another thread. You need to read one repo

bookmark it. the model gets the headlines, the harness does the work