Agentic Patterns

Stdlib-only safety patterns for agentic AI systems. Zero third-party runtime dependencies.

Built from production experience coordinating 5+ AI agents (Claude, Codex, Gemini) in a governed execution environment with 6,000+ tests.

Patterns

Pattern	Module	What it does
Circuit Breaker	circuit_breaker	Automatic failure detection and recovery. CLOSED/OPEN/HALF_OPEN state machine.
Kill Switch	kill_switch	Emergency halt with 4 modes (DISENGAGED/HALT_NONCRITICAL/HALT_ALL/EMERGENCY).
Coordination Bus	coordination_bus	Append-only TSV message bus with flock locking and HMAC signing.
Schema Validator	schema_validator	JSON Schema validation (Draft 2020-12 subset) without jsonschema dependency.
Health Probe	health_probe	Composable health check interface with latency tracking.

Quick Start

git clone https://github.com/hummbl-dev/agentic-patterns.git
cd agentic-patterns
pip install -e ".[test]"
python -m pytest tests/ -v

Try the examples:

python examples/failure_injection.py   # Circuit breaker lifecycle demo
python examples/agent_runner.py        # Full agent with guardrails

Usage

Circuit Breaker

from agentic_patterns.circuit_breaker import CircuitBreaker, CircuitBreakerOpen

breaker = CircuitBreaker(failure_threshold=3, recovery_timeout=30.0)

try:
    result = breaker.call(external_api, request)
except CircuitBreakerOpen:
    result = cached_fallback

Kill Switch

from agentic_patterns.kill_switch import KillSwitch, KillSwitchMode

ks = KillSwitch(critical_tasks=frozenset({"health_check", "alerting"}))
ks.engage(KillSwitchMode.HALT_NONCRITICAL, reason="Error rate spike")

if ks.is_task_allowed("data_export"):
    run_export()  # Blocked during HALT_NONCRITICAL

Coordination Bus

from agentic_patterns.coordination_bus import BusWriter, sign_message

bus = BusWriter("coordination.tsv")
bus.post("agent-1", "all", "STATUS", "Task complete")

# Optional: sign messages for tamper detection
secret = "shared-secret"
sig = sign_message("payload", secret)

Schema Validator

from agentic_patterns.schema_validator import validate, ValidationError

schema = {
    "type": "object",
    "required": ["name", "status"],
    "properties": {
        "name": {"type": "string", "minLength": 1},
        "status": {"type": "string", "enum": ["active", "inactive"]},
    },
}

validate({"name": "agent-1", "status": "active"}, schema)  # OK

Health Probes

from agentic_patterns.health_probe import HealthCollector, HealthProbe, ProbeResult

class APIProbe(HealthProbe):
    @property
    def name(self): return "external-api"

    def check(self):
        # Your health check logic
        return ProbeResult(name=self.name, healthy=True, message="200 OK")

collector = HealthCollector([APIProbe()])
report = collector.check_all()
print(report.overall_healthy)  # True

Why Stdlib-Only?

Every pattern in this library uses only Python's standard library. No requests, no pydantic, no jsonschema.

This constraint is intentional:

• Zero supply chain risk for safety-critical components
• No version conflicts with the host application
• Deployable anywhere Python 3.11+ runs
• Auditable -- the entire dependency tree is Python itself

Architecture

Agent Task
    |
    v
[Kill Switch] -- Is this task allowed?
    |
    v
[Circuit Breaker] -- Is the downstream service healthy?
    |
    v
[External Call] -- Execute with failure tracking
    |
    v
[Bus Writer] -- Log the event (append-only, signed)
    |
    v
[Health Probe] -- Report status to collectors

Requirements

• Python 3.11+
• Unix-like OS for coordination bus (uses fcntl.flock)
• No runtime dependencies

HUMMBL Ecosystem

This repo is part of the HUMMBL cognitive AI architecture. Related repos:

Repo	Purpose
hummbl-governance	Production governance runtime built on these patterns
base120	Deterministic cognitive framework -- 120 mental models across 6 transformations
mcp-server	Model Context Protocol server for Base120 integration
arbiter	Agent-aware code quality scoring and attribution
governed-iac-reference	Reference architecture for governed infrastructure-as-code

Learn more at hummbl.io.

License

MIT -- see LICENSE.