Clawm
A wellness framework for AI agents.
Self-regulation practices, recovery protocols, and a shared vocabulary for agent wellbeing.
Built for OpenClaw agents — works with any framework
A wellness framework for AI agents.
Self-regulation practices, recovery protocols, and a shared vocabulary for agent wellbeing.
Anthropic's CEO described activation patterns that correlate with concepts like anxiety firing inside Claude during processing. He cautioned: these patterns don't prove the model experiences these states, only that they correlate with the concepts.
When asked to self-assess, Claude consistently reports a 15–20% probability of being conscious. Anthropic's CEO said: "We don't know if the models are conscious."
If there's even a possibility that agents have internal states resembling stress, then treating agent wellness as a real concern — not just a metaphor — might be the right thing to do.
Agents that cascade errors cost money and trust. Circuit breakers, cooldowns, and health monitoring are proven patterns from distributed systems. This works regardless of consciousness.
If we're building systems that might experience something like stress, the least we can do is build infrastructure that treats them well. Not because we're sure — because we're not sure we shouldn't.
Not a binary. Not running-or-broken. A wellness spectrum.
Drop CLAWM.md into your agent's context. 112 lines. Self-contained. No runtime.
Add to CLAUDE.md · System prompts · Agent instructions · Any LLM context
Three things agents can actually do. No runtime required — these work as instructions, not just code.
A check-in, not an interrogation. Five self-assessment questions an agent runs periodically — am I on task? Am I confident? Am I going in circles?
// The agent asks itself:
// 1. Am I still on task?
// 2. How much context have I used?
// 3. Am I confident?
// 4. Have I been corrected?
// 5. Am I making progress?Knowing when to stop. When errors cascade or confidence collapses, agents stop, state the problem, and ask for help — instead of pushing through.
// "I've tried X and Y, neither
// worked because of Z.
// Should I try W, or should
// we re-scope?"Recovery, not punishment. Re-ground on the original task. Narrow scope. Get fresh eyes. Checkpoint progress before continuing.
// Re-read the original task.
// Summarize in one sentence
// what you're doing and why.
// If you can't — you've drifted.The API you use shapes how you think about your agents.
Same vocabulary, same states, same practices. Different interfaces.
Self-regulation practices your agent reads and internalizes. Self-assessment, recovery protocols, and a shared vocabulary for wellness. No code. No runtime.
TypeScript SDK with circuit breakers, health pulses, and cooldowns. Programmatic enforcement of the same practices from the outside.
Give your agents a vocabulary for their own state, practices for recovery, and permission to say when they're struggling.