Provides a control-systems metalayer for developing LLM-as-controller agents with safety shields, typed schemas, and multi-rate loop hierarchies.
- GitHub stars
- 5
- Files
- 21
- Last commit
- Nov 11, 2026
- License
- MIT
- Allowed tools
- —
What does the agentic-control-kernel skill do?
agentic-control-kernel provides a control-systems metalayer for developing LLM-as-controller agents with safety shields, typed schemas, and multi-rate loop hierarchies.
Summary generated from the skill's SKILL.md.
About the agentic-control-kernel skill
Developing LLM-as-controller agents with formal safety constraints and hierarchical loops is the primary function of this metalayer. It provides a unified framework for engineers to bootstrap new repositories with typed schemas for states, actions, and traces. By implementing safety shields and plant interfaces, it ensures that an agent's decision-making remains within certified boundaries.
Developers use this skill to establish a multi-rate hierarchy where high-level supervisory planning occurs at the second level while deterministic servo loops run in milliseconds. The system relies on a variety of integrated subsystems including the symphony-orchestrator for dispatch, the knowledge-graph-memory for episodic context, and the autoany ecosystem for continuous controller improvement. It produces essential configuration files like policy.yaml and plant.yaml to define setpoints and system interfaces. This approach differs from standard agent setups by treating the LLM as a producer of typed control directives rather than raw, unconstrained actions.
What an agent can do with it
- 1Bootstrap a new agentic projectInitialize a repository with governance primitives, JSON schemas, and control loop definitions using a single command.
- 2Automate controller optimizationSet up an EGRI loop to continuously improve controller artifacts through automated evaluation.
Written from the skill's SKILL.md. Every use case is checked against a quote from it.
How to install the agentic-control-kernel skill
Each command copies the skill folder into the client's skills directory. Restart the client or start a new session to load it.Install the agentic-control-kernel skill in Claude Code
~/.claude/skills/agentic-control-kernel/SKILL.md
$ npx degit broomva/skills/skills/governance/agentic-control-kernel ~/.claude/skills/agentic-control-kernelInstall the agentic-control-kernel skill in Claude Code (project)
.claude/skills/agentic-control-kernel/SKILL.md
$ npx degit broomva/skills/skills/governance/agentic-control-kernel .claude/skills/agentic-control-kernelInstall the agentic-control-kernel skill in Codex
~/.agents/skills/agentic-control-kernel/SKILL.md
$ npx degit broomva/skills/skills/governance/agentic-control-kernel ~/.agents/skills/agentic-control-kernelInstall the agentic-control-kernel skill in Gemini CLI
~/.gemini/skills/agentic-control-kernel/SKILL.md
$ npx degit broomva/skills/skills/governance/agentic-control-kernel ~/.gemini/skills/agentic-control-kernelInstall the agentic-control-kernel skill in Claude.ai
npx degit broomva/skills/skills/governance/agentic-control-kernel agentic-control-kernel && zip -r agentic-control-kernel.zip agentic-control-kernel # claude.ai → Customize → Skills → + → Create skill → Upload a skill → agentic-control-kernel.zip
Automated checks
Ran Nov 15, 2026Full report →Passed our automated checks. That is not a security audit: read the scripts before you let an agent run them. How checks work
- Source repository linked and reachablebroomva/skills/skills/governance/agentic-control-kernel is public and reachablegithub.com/broomva/skills/tree/HEAD/skills/governance/agenti ↗Pass
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit 2026-11-11github.com/broomva/skills/commits ↗Pass
- No download-and-run or credential theft in scripts3 scripts read; no download-and-run, remote eval, base64 execution or credential uploadPass
- No hidden instructions in SKILL.mdNo override phrases, safeguard switches, credential requests or hidden charactersgithub.com/broomva/skills/blob/HEAD/skills/governance/agenti ↗Pass
- Pre-approved tools are scopedNo pre-approved tools; the agent asks before using onegithub.com/broomva/skills/blob/HEAD/skills/governance/agenti ↗Pass
- SKILL.md follows the Agent Skills formatValid name and description; says when to use it; 145 linesgithub.com/broomva/skills/blob/HEAD/skills/governance/agenti ↗Pass
SKILL.md
View on GitHubAgentic Control Kernel
A purely knowledge-based metalayer that unifies six subsystems into a single installable skill for any project:
| Layer | Source / Crates | Role |
|---|---|---|
| Governance | control-metalayer-loop | Setpoints, sensors, gates, policy, profiles |
| Improvement | autoany_core + autoany-aios + autoany-lago |
EGRI microkernel, Arcan execution, Lago ledger |
| Orchestration | symphony-orchestrator + symphony-arcan |
Poll/dispatch/worker/reconcile via Arcan HTTP |
| Runtime | Life (arcan, lago, autonomic, praxis, spaces) |
Agent sessions, event journal, homeostasis, networking |
| Protocol | aios-protocol |
Canonical types — shared vocabulary across all crates |
| Episodic Memory | knowledge-graph-memory | Conversation logs -> Obsidian bridge |
| Consciousness | agent-consciousness | Three-substrate persistent context |
| QA/Actuation | gstack | Headless browser, workflow skills |
| Control Kernel | this skill | Plant interface, safety shields, typed schemas, multi-rate hierarchy |
Core Law
Do not grant an agent more mutation freedom than your evaluator can reliably judge. In control terms: do not let the LLM's action space exceed what your runtime monitors, safety filters, and evaluators can certify.
Quick Start
1. Bootstrap a project
python3 scripts/control_kernel_init.py <repo-path> [--profile governed] [--runtime arcan] [--ledger lago]
This installs into the target repo:
.control/policy.yaml— control-systems-aware setpointsschemas/— state, action, trace, evaluator JSON schemasMETALAYER.md— control loop definition with plant/shield/estimator sections- Harness gates wired to
make smoke,make check,make control-audit
2. Define the plant interface
Edit .control/plant.yaml with typed state and action schemas for your system.
See references/plant-interface.md for the full API spec.
3. Wire safety shields
See references/safety-shields.md for CBF-QP patterns, policy gates, and containment invariants.
4. Set up EGRI for controller improvement
Use the problem-spec template in assets/templates/problem-spec.control.yaml
to define an autoany loop over your controller artifacts.
See references/egri-for-controllers.md.
Architecture Overview
The LLM emits typed control directives θ_t — not raw actuations u_t.
Deterministic controller modules execute, safety shields filter, and the runtime
logs traces to an append-only ledger.
Plant → observe() → Runtime → update estimator → b_t
→ LLM Agent: request decision(b_t) → θ_t (typed directive)
→ Controller: propose(b_t, θ_t) → proposed u_t
→ Safety Shield: filter(u_t, b_t) → safe u_t + certificate
→ Plant: apply(safe u_t) → result
→ Evaluator/Ledger: append trace + score
See references/architecture.md for the full 5-layer diagram.
Multi-Rate Hierarchy
| Loop | Cadence | LLM here? | What runs |
|---|---|---|---|
| Servo | ms | No | PID, state feedback, deterministic |
| Constrained execution | 10-100ms | No (param updates only) | MPC/CBF-QP solvers |
| Supervisory planning | seconds | Yes | Goal setting, mode switching, tool selection |
| Auto-tuning (EGRI) | minutes-days | Yes | Controller synthesis, model learning |
See references/multi-rate-hierarchy.md.
LLM Roles in the Control Stack
| Role | Outputs | When to use |
|---|---|---|
| Supervisory controller | setpoints, mode switches, constraints | Default — long-horizon reasoning |
| Meta-controller | tool/module selection, identification triggers | Modular systems with multiple controllers |
| Controller synthesizer | code, configs, tests | Offline — gated by harness CI |
| EGRI loop compiler | problem-spec, evaluator design, promotion rules | Continuous improvement cycles |
See references/architecture.md for the full role table.
Reference Guide
- architecture.md — 5-layer stack, realized crate graph, control-flow diagram, component mapping
- integration-map.md — Adapter crate boundary map, configuration, direction rule
- plant-interface.md — Plant/Estimator/Controller/Shield/Evaluator API specs
- safety-shields.md — CBF-QP, policy gates, containment, failure modes
- multi-rate-hierarchy.md — Loop rates, LLM placement, heuristics
- world-models.md — Koopman, DeePC, digital twins, learned dynamics
- egri-for-controllers.md — Autoany applied to controller optimization
- orchestration-patterns.md — Symphony daemon patterns for multi-agent dispatch
- consciousness-stack.md — Memory/knowledge/episodic integration
- failure-modes.md — Mitigations catalog for LLM-in-the-loop control
- deep-research-report.md — Original research report and project plan: formal control theory, literature survey, prototype roadmap
Schemas
JSON Schemas in schemas/ enforce typed interfaces:
state.schema.json— Plant/belief stateaction.schema.json— Control directives (θ_t)trace.schema.json— Ledger entries (autoany-compatible)evaluator.schema.json— Score vectors, promotion decisionsegri-event.schema.json— EGRI trial events for Lago persistence via EventKind::Custom
Existing Skill Dependencies
This skill synthesizes and references (does not duplicate) these existing skills:
- control-metalayer-loop — Use for
.control/bootstrapping and governance primitives - autoany — EGRI loop execution via
autoany-aios(Arcan sessions) andautoany-lago(Lago ledger) - symphony — Orchestration dispatch via
symphony-arcan(Arcan HTTP runtime) - life —
arcan(agent sessions),lago(event journal),autonomic(homeostasis),spaces(networking) - aios-protocol — Canonical types shared across all adapter crates
- agent-consciousness — Use for consciousness stack setup
- knowledge-graph-memory — Use for conversation bridge to Obsidian
- gstack — Use for QA actuation via headless browser
Files in this skill
21agentic-control-kernel skill questions
How do I install the agentic-control-kernel skill in Claude Code?
Run npx degit broomva/skills/skills/governance/agentic-control-kernel ~/.claude/skills/agentic-control-kernel. Claude Code loads skills from ~/.claude/skills (all projects) or .claude/skills (one project) and uses a skill when its description matches the task.
Which clients can use the agentic-control-kernel skill?
Any client that reads the Agent Skills format: Claude Code, Codex, Gemini CLI and claude.ai (upload as a ZIP). Scripts in the skill run only where the client allows code execution.
How do I bootstrap a new project with the control kernel?
You can bootstrap a project by running the control_kernel_init.py script with your repository path and optional profiles for governance or runtime.
What is the core principle regarding agent mutation freedom?
The core law states that you should not grant an agent more mutation freedom than your evaluator can reliably judge, ensuring the action space does not exceed what monitors and safety filters can certify.