Judgement for long-running agents.
SNAXK is a research skill for OpenClaw testing whether a lightweight control layer can help long-running agents act with better judgement, better boundaries, and clearer review.
Current research milestone: SNAXK 0.10.8.
Currently being tested first as a skill for OpenClaw. Closed-source for now, and still in active research rollout.
Agents can produce output before they know how to judge it.
Most agent systems are already good at taking a task and producing an answer. The real gap is judgement.
  • when to slow down
  • when a request is risky
  • what is appropriate to share
  • how to adapt to the person they are dealing with
  • how to improve without drifting out of control
That is the space SNAXK is operating in.
We are testing a simple hypothesis.
We think agents behave better when they do not jump straight from input to action.
Instead, we think they should:
01
Run a fast heuristic pass
over messages, events, and signals
02
Decide whether anything meaningful
is happening
03
Route triggered cases
through bounded judgement and guardrails
04
Reflect overnight
05
Only carry forward changes
that survive review
The practical question is:
  • is this safer
  • is it more understandable to humans
  • is it better than simpler alternatives
  • can we measure whether the system is becoming too cautious, too permissive, or more trustworthy over time
  • can the control layer name common agentic attack pressures explicitly without pretending prompts alone are a security boundary
  • can nightly review recommend rule and threshold changes without turning into uncontrolled self-modification
The problem is not output. The problem is judgement.
Long-running agents do not usually fail because they cannot generate text.
They fail because they:
Over-share
Over-react
Drift out of alignment
Lose track of trust and disclosure boundaries
Accumulate too much stale context
Change in ways that are hard to supervise
That is the space SNAXK is operating in.
What SNAXK adds.
It sits on top of an existing host system and adds:
Fast first-pass signals
Bounded judgement
Clearer response control
Compact awareness
Bounded group-chat triage
Review and audit surfaces
Clearer governance and rollout posture
Security-aware boundary thinking
The first live proving ground is OpenClaw.
How it works.
01
Signals
A fast pass over signals such as urgency, distrust, surprise, pressure, and fear. Hooks provide bounded intake from the host.
02
Judgement
head = evidence. heart = dignity and trust impact. gut = anomaly and pre-harm sensing. spine = hard boundary.
03
Response
Decide whether to proceed, narrow, verify, summarise, refuse, or escalate. Conversation posture is inspectable before reply.
04
Review
Nightly reflection generates bounded proposals. Morning review decides what carries forward. Nothing changes without human sign-off.
Research progress
The work is still active research, but the milestones are now real and inspectable.
Each release is aimed at better judgement, better reviewability, and stronger human trust.
0.5.0
Stitched judgment audit
0.6.0
Verdicts, outcomes, and judgment metrics
0.7.1
Local replay
0.8.0
Stronger evaluation
0.9.0
Clearer runtime legibility
0.10.2
Bounded group-chat triage
0.10.3
Cleaner persona boundary
0.10.7
Explicit hooks, reviewable conversation state, and bounded recommendations under review
0.10.8
Clearer governance defaults and cleaner rollout readiness
This is a research log, not a product changelog. Milestones reflect what has been built and tested, not what is finished.
Better judgement makes agents more usable.
The point is not to make agents sound more human.
The point is to make them:
  • safer to deploy
  • easier to supervise
  • easier for people to work with
  • more stable over long periods
We are also testing whether this is actually better than simpler alternatives:
  • direct if/then routing
  • prompt-only interpretation
  • hidden autonomous adjustment
Why this approach exists.
We become just by doing just acts. — Aristotle
That is close to the spirit of SNAXK.
Judgement does not come from one prompt or one policy document. It comes from repeated practice, feedback, review, and correction. That applies even more now that the system can record human reviews and begin measuring whether repeated judgement is actually improving.
All real living is meeting. — Martin Buber
On trust, counterpart, and communication.
Who looks outside, dreams; who looks inside, awakes. — Carl Jung
On reflection, sleep, and review.
This is not a personality skin.
SNAXK is not:
  • a vibe layer
  • a replacement for your agent framework
  • a replacement for risk or security systems
  • an uncontrolled self-improvement loop
It is a control layer for judgement, boundaries, and communication.
It is also not being presented as a finished generally available product. It is a research system under live test.
OpenClaw is the first proving ground.
SNAXK is currently being tested first as a skill for OpenClaw.
Current research milestone: SNAXK 0.10.8
Latest work: clearer governance defaults and cleaner rollout readiness
Why SNAXK?
We call it SNAXK because it is designed to take small, bounded bites of signals, messages, and context.
Instead of one large opaque judgement, it breaks interpretation into smaller, reviewable steps.
From task execution to operational judgement.
SNAXK helps agents know:
What is happening
Who they are dealing with
What is safe to say
When to proceed
When to stop
How to improve without losing control
What is safe to name
Security-aware spine posture for common agentic attack pressures
What can be proposed
Bounded rule and threshold recommendations under review before anything carries forward
Join the research preview
We are looking for research-minded operators and experimental teams running long-lived agents. If you are interested in safety, legibility, and supervised AI control — and willing to share structured feedback from real usage — we want to hear from you.
This is not a general availability launch. Participation is limited and feedback-driven.
Try a limited deployment
Request access to run SNAXK in a real environment and report back what you observe.
Send structured feedback
Already running something relevant? Send us your learnings directly — anonymously or attributed.
What to include in feedback
What happened — describe the agent behaviour you observed
What worked — where the control layer behaved as expected
What felt off — where something seemed wrong, noisy, or misaligned
What the system misunderstood — signals it read incorrectly or missed entirely
What you think should change — your interpretation of what the architecture got wrong
Whether the system felt too cautious, too permissive, or about right — help us test whether judgment quality is improving over time
Whether the replay, scoring, and review loop is surfacing the right changes — help us test whether local replay is catching what live nights miss
Current research milestone: SNAXK 0.10.7
Latest work: security-aware spine posture, explicit hook contracts, reviewable conversation state, and bounded rule or threshold recommendations under review
Feedback can be anonymous or attributed. Avoid private or sensitive personal information. Feedback is used to improve the architecture — not for marketing or public attribution without consent.