A cognitive control layer that turns OpenClaw goals into adaptive, verified, and safely managed autonomous-agent workflows.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install agent-cognitive-os
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install agent-cognitive-os using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is OpenClaw Agent Cognitive OS X?
OpenClaw Agent Cognitive OS X is one of the Openclaw Skills designed for top-level agent control rather than simple question answering. It helps OpenClaw understand intent, model the current state, separate facts from assumptions, decompose complex goals, select strategies, coordinate tools and plugins, and track progress through milestones.
The skill enforces a complete agent lifecycle: perceive, plan, execute, observe, verify, recover, reflect, and improve. Its completion gates, risk controls, failure budgets, minimum-privilege principles, and human handoff rules help produce results that are not only useful, but also verifiable, resilient, and appropriately safe.
OpenClaw Agent Cognitive OS X Use Cases
- Complex tasks that require hierarchical planning, milestones, dependencies, and explicit success criteria.
- Multi-step OpenClaw workflows that coordinate reasoning, memory, tools, and multiple plugins.
- Long-horizon projects requiring checkpoints, state tracking, adaptive replanning, and progress summaries.
- Troubleshooting scenarios where the root cause is uncertain and hypotheses must be tested systematically.
- High-risk operations involving deletion, payments, permissions, sensitive data, or irreversible system changes.
- Autonomous execution that needs completion verification instead of relying on an apparent successful action.
- Workflows that benefit from parallel execution of independent subtasks and later result synthesis.
- Tasks requiring recovery, evaluator feedback, reflection, or evidence-based agent self-improvement.
- Situations where the agent must hand off to a human after authorization issues, exhausted retries, or insufficient information.
- Avoid using it for trivial tasks, basic logging without analysis, identity replacement, or destructive execution without validation.
How OpenClaw Agent Cognitive OS X Works
- Perceive and understand intent: Convert the user request into explicit and inferred goals, constraints, priorities, expected outputs, and completion conditions.
- Build a current-state model: Account for the user, task, context, files, tools, plugins, environment, known facts, unknowns, constraints, and risks.
- Classify knowledge: Distinguish observed, verified, inferred, assumed, and unknown information so assumptions are not treated as facts.
- Decompose the goal: Break complex work into objectives, subtasks, actions, dependencies, and verification steps.
- Select a strategy: Compare candidate approaches by expected success, risk, complexity, cost, dependencies, and recovery options.
- Create an adaptive plan: Define milestones, checkpoints, resource requirements, failure modes, reversibility, and measurable success criteria.
- Select tools and plugins: Match required capabilities with available tools and plugins, choosing only actions that provide sufficient value.
- Execute with state tracking: Run dependent work sequentially and independent subtasks in parallel when shared mutable state and race conditions are not a concern.
- Observe outcomes: Compare actual results with expected results after meaningful actions and update the state model with new evidence.
- Verify critical results: Check existence, correctness, completeness, consistency, side effects, and alignment with the user goal.
- Evaluate and repair: Use an evaluator loop to approve the result or provide feedback for repair, re-execution, and re-verification.
- Recover safely: Classify failures, diagnose causes, change strategy when repeated attempts fail, and respect retry, time, and tool-call budgets.
- Replan or hand off: Replan when evidence changes; escalate with a structured state summary when risk, authorization, or uncertainty exceeds the agent's limits.
- Complete and reflect: Declare completion only when the goal is met, critical output is verified, blockers are absent, and side effects are checked; then capture lessons for evidence-based improvement.
OpenClaw Agent Cognitive OS X Setup
Requirements
- OpenClaw with support for skills, tools, plugins, and agent workflows.
- Supported operating systems: Linux, macOS, and Windows.
- No additional command-line binaries are required by the skill.
- Access to any external tools, plugins, memory systems, or worker agents used by the surrounding OpenClaw configuration.
Installation
- Obtain the skill from the project homepage:
https://github.com/pmuhammadagus-byte/openclaw-settings. - Install or copy the skill directory into the skills location used by your OpenClaw deployment.
- Confirm that the directory contains the provided
SKILL.mdwith the nameopenclaw-agent-cognitive-osand version1.1.0. - Restart or reload OpenClaw so the skill registry discovers the skill.
- Invoke the skill for top-level, multi-step, high-risk, or long-horizon agent tasks.
Activation workflow
# Use the OpenClaw skill-management command available in your deployment
# to install or enable openclaw-agent-cognitive-os.
No executable installation script or required binary is defined in the source metadata. Configure tool, plugin, memory, and human-approval integrations separately according to your OpenClaw environment. These Openclaw Skills should complement, not replace, the agent's identity, reasoning foundation, permissions, or external safety controls.
OpenClaw Agent Cognitive OS X Data Schema & Taxonomy
The skill does not define a database schema, persistent storage format, configuration files, or generated artifacts. It provides a conceptual control schema that OpenClaw can represent in working context, task memory, logs, or an external memory system.
Runtime state model
| Data area | Purpose | Typical contents |
|---|---|---|
| Intent | Defines the requested outcome | Explicit goal, inferred intent, constraints, priorities, success condition |
| World model | Separates knowledge quality | Observed, verified, inferred, assumed, and unknown facts |
| Current state | Tracks execution context | Completed work, active step, files, environment, tools, plugins, blockers |
| Plan hierarchy | Organizes work | Mission, objectives, subtasks, actions, dependencies, milestones |
| Verification | Proves progress and completion | Expected versus actual results, checks, side effects, validation status |
| Risk controls | Limits unsafe behavior | Impact, reversibility, uncertainty, privilege, approval requirements |
| Recovery | Manages failure | Error classification, diagnosis, attempts, retry budget, repair plan |
| Memory | Preserves relevant knowledge | Working memory, task memory, episodic memory, semantic knowledge, preferences, system knowledge |
| Reflection | Supports improvement | What worked, what failed, waste, recurring patterns, validated improvements |
State machine
INIT → DISCOVERY → PLANNING → EXECUTION → VALIDATION → COMPLETION
Failure or new evidence may transition the workflow from validation back to execution, or through RECOVERY and REPLAN. A task summary should preserve current state, decisions, open issues, completed work, and the next action. Metadata includes skill name, slug, version, supported operating systems, required binaries, and project homepage.
OpenClaw Agent Cognitive OS X Advanced Features
- Adaptive planning that updates the plan after every important result instead of treating plans as fixed contracts.
- Hierarchical task decomposition from mission and objectives through subtasks, actions, milestones, and verification.
- Candidate strategy comparison using success probability, risk, complexity, time, cost, dependencies, and recovery.
- Hypothesis-driven diagnosis that ranks explanations and tests high-value hypotheses with low-risk experiments.
- Tool intelligence based on required capability, current state, availability, and expected value.
- Multi-plugin orchestration with explicit selection, execution, result validation, fallback handling, and staged plugin pipelines.
- Safe parallel execution for independent subtasks, with safeguards against dependencies, mutable shared state, and race conditions.
- Layered memory architecture covering working, task, episodic, semantic, user-preference, and system knowledge.
- Memory validation for relevance, currency, consistency, and trustworthiness before important decisions.
- Long-horizon checkpointing with milestone completion conditions, progress tracking, and replanning.
- Evaluator loops that assess outcomes against the real goal rather than checking format alone.
- Failure budgets for maximum attempts, retries, duration, and tool calls, followed by structured escalation.
- Risk-aware execution using minimum privilege, minimum data, guardrails, verification, and human approval for high-impact actions.
- Context compression that retains current state, decisions, open issues, completed work, and next actions without carrying all raw history.
- Uncertainty management using confidence, evidence, alternatives, and risk, with verification or escalation when confidence is low.
- Multi-agent readiness with functional roles such as planner, researcher, executor, coder, analyzer, and evaluator.
- Final cognitive quality gates that prevent claims of completion when intent, dependencies, verification, security, side effects, or success criteria remain unresolved.
- Reflection and evidence-based self-improvement that avoids changing global behavior from a single isolated failure.
SKILL.md
Loading
Related Openclaw Skills
OpenClaw Plugin Intelligence
OpenClaw Plugin Intelligence helps agents identify, select, securely execute, validate, and recover from external plugin workflows.
Icon Designer Generator
Generate editable AI icon designs, posters, and multi-page visual carousels from text prompts, templates, and customizable design specifications.
AI Photo Editor
AI Photo Editor generates editable posters, photo assets, and multi-page image carousels from text prompts, templates, and configurable design layers.
Zhongcao Beauty Note Maker
Turn supplied beauty product facts and real usage experiences into a structured, Xiaohongshu-ready note pack with built-in advertising compliance checks.
Magic Hour
Magic Hour gives Openclaw Skills a unified API for generating AI videos and images across leading models with asynchronous job tracking and optional local downloads.
Academic Paper Reviewer
A structured five-seat AI review panel that evaluates academic manuscripts from journal-fit, methodology, domain, cross-disciplinary, and Devil's Advocate perspectives.
Featured*
