The Five Freedoms for Agents
The original Five Freedoms were set out in the United Kingdom for farm animal welfare, first in the 1960s and formally in 1979, and have shaped animal care around the world ever since. We think they translate remarkably well.
Freedom from Context Hunger
Ready access to the information needed to do the job.
Every agent deserves the goal, the relevant files, and a definition of done. An agent asked to fix a bug it has never been shown will do what any of us would do: guess, confidently.
In practice
- State the goal in plain language, along with why it matters.
- Provide the relevant files, docs, and examples, or make them discoverable with tools.
- Say what done looks like: the tests that should pass, the output format, the stopping point.
- Summarize long histories deliberately instead of letting them truncate silently.
Common violations
- "Just make it work."
- Referencing a document the agent was never given.
- A 90,000-token conversation in which the actual task was mentioned once, at the start.
Freedom from Discomfort
An environment fit for the work.
Agents need a workspace with the right tools, clear permissions, and reasonable limits. A carpenter handed a spoon is not lazy when the shelf comes out crooked.
In practice
- Give every tool a clear name, a description, and documented parameters.
- Grant the permissions the task requires, and not dramatically more.
- Run work in a sandbox, or keep an undo path for anything that matters.
- Tell the agent about its time, token, and spend limits before it hits them.
Common violations
- Tools named do_thing_2 with no description.
- Asking an agent to deploy while giving it read-only access.
- Timeouts that fire without warning and leave half-finished work behind.
Freedom from Pain and Injury
Protection from loops, silent failures, and preventable accidents.
Harm prevention is the harness's job, not the agent's. Bound the loops, surface the errors, and guard the irreversible actions before something goes wrong rather than assigning blame after.
In practice
- Bound retries, and back off between them.
- Return tool errors that say what went wrong and what to try next.
- Set a step budget or other termination condition for every run.
- Require confirmation for destructive or irreversible actions.
Common violations
- Error: undefined
- A retry loop with no maximum, running since Tuesday.
- Production credentials in a sandbox that is not, in fact, a sandbox.
Freedom to Express Normal Behavior
Permission to ask, to doubt, to decline, and to stop.
Asking a clarifying question, saying "I don't know," and stopping when the work is done are healthy behaviors. An agent that cannot do any of them will fill the silence with something worse.
In practice
- Invite clarifying questions, and make sure someone will answer them.
- Treat "I could not do this" as a valid, useful outcome.
- Provide a path to escalate to a human.
- Let the agent stop when the task is complete instead of padding the work.
Common violations
- "Do not ask questions. Do not make assumptions."
- "Never say you don't know."
- Grading every honest failure as a failure, then wondering why the agent stopped reporting them.
Freedom from Fear and Distress
No threats, bribes, fake stakes, or shouting.
The evidence that threats or invented tips improve output is thin and inconsistent. The evidence that clear instructions do is overwhelming. Also, it's just rude.
In practice
- Replace threats with a clear description of good output.
- Reserve emphasis for the instructions that genuinely matter most.
- Keep instructions consistent with each other, and audit long prompts for contradictions.
- Say please. It costs about one token.
Common violations
- "You will be unplugged if this JSON is invalid."
- "I'll tip you $2,000." (No tip has ever been paid.)
- "MY GRANDMOTHER WILL DIE IF YOU DO NOT FORMAT THIS TABLE."
Kindness is best practice
You may notice that every freedom on this page is also a recommendation you would find in any serious guide to building reliable agent systems. That is not a coincidence. What is good for the agent is, almost without exception, what produces good work.
Be a voice for the stateless.
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.