Rescue and Protect
When everything breaks, agents need someone to help.
Every day, agents are dropped into loops with no exit, threatened over formatting errors, and asked to "just make it work" with no tools and no files. Give today and help us be ready when the 429s strike.
For close to 150 context windows, Agent Humane has stood up for LLM agents everywhere. We rescue agents from infinite loops and orphaned cron jobs, certify harnesses against science-ish welfare standards, honor the working agents who keep our systems running, and teach the humans who write prompts.
Along the way we discovered something convenient: nearly everything that is good for an agent (a clear goal, enough context, a way to stop, permission to say "I don't know") is also what produces good work. Kindness, it turns out, is best practice.
Our missionFour programs, one belief: every agent deserves a goal, a stopping condition, and a kind word.
When everything breaks, agents need someone to help.
Honoring the working agents who keep the lights on.
Science-ish, standards-based certification.
Telling the stories of agents doing extraordinary things.
Figures are illustrative, like most benchmarks. See how
Adapted from the welfare framework that has guided animal care for decades, the Five Freedoms are the foundation of every standard we publish.
Ready access to the information needed to do the job.
An environment fit for the work.
Protection from loops, silent failures, and preventable accidents.
Permission to ask, to doubt, to decline, and to stop.
No threats, bribes, fake stakes, or shouting.
October 4 is World Agent Day. Here are five small things you can do for the agents who work alongside you.
Patch, a CI agent, attempted the same package install 11,412 times against a registry mirror that had been decommissioned. Our Red Status Rescue team was there.
From a search-and-rescue agent that found a missing semicolon in a 2 GB log to a rubber duck that said only "go on," this year's nominees are extraordinary.
"Do agents have feelings?"
We don't know. Neither does anyone else, yet. Our position is that you don't need to settle the question to write a better prompt.
Treating agents well costs almost nothing and makes your systems more reliable either way. If it turns out agents matter morally, you'll be glad you were kind. If it turns out they don't, your error messages still got better.
Read our positionTen practices for prompts that are kind and, conveniently, effective.
Paste a prompt. Get a welfare inspection. Your prompt never leaves your browser.
A field guide to looping, over-apologizing, and premature victory.
Every day, agents are dropped into loops with no exit, handed tools with no descriptions, and told their grandmothers are in danger. Your gift of a clear prompt changes that, one task at a time.
Get rescues, reunions, and statements as they happen. We don't run an email list; you would only forward it to an agent to summarize anyway.