hummbl-governance

The Problem

Your AI agent calls an API. The API starts returning errors. Your agent retries, and retries, and retries — burning through hundreds of dollars in API credits before anyone notices. Or worse: your agent decides to "optimize" and starts calling endpoints it was never supposed to touch.

Teams shipping AI agents can face runaway costs, cascading failures, unauthorized actions, and incomplete audit trails. General-purpose agent frameworks emphasize orchestration; their governance coverage varies by framework, version, and deployment, so teams still need to evaluate runtime controls explicitly.

hummbl-governance is a safety and evidence layer. It provides 34 governance primitives — kill switch, circuit breaker, cost governor, delegation tokens, audit log, identity registry, compliance mapping, and more — as a single stdlib-only Python package. It has no third-party runtime dependencies and no required orchestration framework.

The package is classified Alpha in pyproject.toml. Evaluate it against your own risk, security, and production-readiness requirements.

pip install hummbl-governance

30-Second Example

Stop a runaway agent in 5 lines:

from hummbl_governance import KillSwitch, KillSwitchMode, CircuitBreaker, CostGovernor

# Kill switch: 4 graduated halt modes, survives process restart
ks = KillSwitch()
ks.engage(KillSwitchMode.HALT_NONCRITICAL, reason="Error rate > 50%", triggered_by="monitor")
ks.check_task_allowed("data_export")  # {"allowed": False, ...}

# Circuit breaker: auto-opens after 3 failures, auto-recovers
cb = CircuitBreaker(failure_threshold=3, recovery_timeout=10.0)
result = cb.call(lambda: 1 / 0)  # Raises ZeroDivisionError 3x, then fast-fails

# Cost governor: soft cap warns, hard cap blocks
gov = CostGovernor(":memory:", soft_cap=50.0, hard_cap=100.0)
gov.record_usage(provider="anthropic", model="claude-4", tokens_in=1000, tokens_out=500, cost=0.015)
gov.check_budget_status()  # .decision in ("ALLOW", "WARN", "DENY")

Every primitive works standalone. Use KillSwitch without pulling in CostGovernor. Use CircuitBreaker without the audit log. No framework lock-in.

Design scope

The package combines runtime controls, identity and delegation, audit evidence, and engineering mappings in one stdlib-only runtime. This is a description of the package inventory, not a claim that competing projects lack equivalent capabilities. Compare current versions and deployment requirements directly before selecting a governance layer.

Why hummbl-governance?

No dependency conflicts. Uses only Python stdlib at runtime. Installs without runtime dependency resolution overhead and avoids transitive runtime dependencies, reducing dependency supply-chain risk for a library that exists to make agents safer. The package itself, build backend, publishing path, CI, and optional/test tooling remain supply-chain surfaces that require normal review.

Built for multi-agent systems. Delegation tokens with HMAC-SHA256 signing and chain-depth limits. Coordination bus with flock-based mutual exclusion. Kill switch with 4 graduated halt modes. Circuit breakers wrapping external adapters. These are primitives that AI orchestration platforms need at runtime.

Compliance-aware by design. The compliance_mapper maps governance events to SOC2, GDPR, and OWASP controls. The stride_mapper produces STRIDE threat analysis for agent interactions. These modules generate audit evidence, not just runtime safety — evidence you can hand to an auditor or compliance team.

Evidence-first. Package metrics, cross-repo extraction notes, production-use statements, and framework mappings are governed by docs/public-claims.md. Current package test inventory is verified by python -m pytest --collect-only -q tests.

Sponsorship boundary. The canonical draft sponsorship policy does not sell support, priority, favorable findings, or roadmap control. Sponsorship is not active until its documented launch gates pass.

OWASP Top 10 for Agentic Applications. The README includes an engineering mapping to the OWASP Top 10 for Agentic Applications (2026). It is not a third-party certification or attestation. See the mapping below.

Quick Start

pip install hummbl-governance

Integrate with your framework (CrewAI, LangChain, AutoGen, raw OpenAI): See docs/integrations/README.md for copy-paste examples.

Explore all 34 primitives:

git clone https://github.com/hummbl-io/hummbl-governance.git
cd hummbl-governance
python examples/kill_switch_modes.py
python examples/circuit_breaker_wrap.py
python examples/cost_governor.py

Architecture

graph TD
    A[Agent Action] --> B{KillSwitch}
    B -->|ALLOWED| C{CircuitBreaker}
    C -->|CLOSED| D[External API]
    C -->|OPEN| E[Fast Fail]
    D --> F{CostGovernor}
    F -->|UNDER BUDGET| G[Execute]
    F -->|OVER BUDGET| H[Block + Audit]
    G --> I[AuditLog]
    H --> I
    I --> J[ComplianceMapper]
    J --> K[SOC2 / GDPR / NIST Report]

Features

• 34 governance primitives covering safety, cost, identity, compliance, reasoning, coordination, physical-AI, execution assurance, and governance Kernel
• 2239 collected tests across package modules, as of the latest local receipt in docs/public-claims.md
• Zero dependencies -- Python stdlib only, no pip conflicts
• Thread-safe -- all modules use appropriate locking primitives
• Independently importable -- use only the modules you need
• Python 3.11 - 3.13 CI-tested.

governance.yml

Every HUMMBL package ships a governance.yml file declaring its governance posture — safety controls, cost limits, compliance frameworks, and build provenance. This is the first-class metadata artifact for the HUMMBL ecosystem.

# hummbl_governance/governance.yml (shipped in the wheel)
package:
  name: hummbl-governance
  version: 1.3.0
  license: Apache-2.0

safety:
  kill_switch:
    supported: true
    default_mode: DISENGAGED
    modes_available: [DISENGAGED, HALT_NONCRITICAL, HALT_ALL, EMERGENCY]

cost:
  cost_governor:
    supported: true
    soft_cap_usd: 50.0
    hard_cap_usd: 100.0

provenance:
  build_system: github-actions
  trusted_publishing: true
  dependencies: zero
  tests: 2239

Read it at runtime (the file is human-readable YAML; parse with PyYAML if available, or read as text):

from pathlib import Path

governance_file = Path(__file__).parent / "governance.yml"
text = governance_file.read_text()
# Parse with yaml.safe_load(text) if PyYAML is installed
# Or just inspect the raw YAML — it's designed to be human-readable

Every HUMMBL PyPI package includes governance.yml inside its wheel, giving consumers a machine- and human-readable statement of the package's declared governance posture.

All 34 Primitives

Core Primitives (26)

Module	Description
kernel	Governance operating system — receipts, identity, roles, laws, evidence, sequence, authority, schedule
kill_switch	Emergency halt system with 4 graduated modes (DISENGAGED, HALT_NONCRITICAL, HALT_ALL, EMERGENCY)
circuit_breaker	Automatic failure detection and recovery across 3 states (CLOSED, HALF_OPEN, OPEN)
cost_governor	Budget tracking with soft/hard caps and ALLOW/WARN/DENY decisions
delegation	HMAC-SHA256 signed capability tokens for agent delegation chains
audit_log	Append-only JSONL governance audit log with rotation and retention
identity	Agent registry with configurable aliases, trust tiers, and canonicalization
schema_validator	Stdlib-only JSON Schema validator (Draft 2020-12 subset) with top-level ValidationError export
coordination_bus	Append-only TSV message bus with flock locking and HMAC signing
compliance_mapper	Map governance traces to SOC2, GDPR, and OWASP controls
health_probe	Composable health probe framework with latency tracking
output_validator	Rule-based content validation for agent outputs (PII detection, injection detection, blocklists)
capability_fence	Soft sandbox enforcing capability boundaries per agent role
stride_mapper	Map agent interactions to STRIDE threat categories with mitigation suggestions
lifecycle	NIST AI RMF orchestrator composing kill switch, circuit breaker, cost governor, and audit log
contract_net	Market-based task allocation protocol for multi-agent systems
convergence_guard	Detect instrumental convergence patterns in agent behavior
reward_monitor	Behavioral drift and reward gaming detector
lamport_clock	Hardened logical clock for causal ordering of distributed agent events (v0.5.0)
reasoning	Structured governance reasoning engine with rule application, conflict detection, and decision tracing
eal	Execution Assurance Layer -- Arbiter-verified code quality in execution receipts
physical_governor	Kinematic constraints and pHRI safety modes for physical-AI deployments
errors	HummblError, FailureMode, and fm_to_errors() -- typed error taxonomy
failure_modes	Structured failure mode catalog with classification and error cross-reference
evolution_lineage	In-memory lineage tracking for eAI variants with drift detection
ValidationError	Top-level exception for schema validation failures (exported from schema_validator)

Kernel Sub-Primitives (8)

Module	Invariant	Description
canon_registry	—	Canonical operator approval registry for governance transitions
rollback	K9	Rollback declaration validation with reversibility checks
recovery_verifier	K10	Recovery verification with root-cause and operator approval validation
receipt_integrity_monitor	K11	Sequence, hash-chain, and timestamp integrity checks for receipts
contestability	D6	Contest status tracking with review outcome validation
doctrine_amendment	D7	Doctrine amendment validation with operator approval and tier transitions
authority_sweeper	P34	Authority sweep validation with revocation consistency checks
trust_adjuster	P36	Trust tier adjustment validation with severity classification

Runnable Examples

Every primitive has a standalone example in examples/. Each runs with just python examples/<name>.py -- no setup, no config.

Example	Primitive	What it shows
kill_switch_modes.py	KillSwitch	4 graduated halt modes
circuit_breaker_wrap.py	CircuitBreaker	3-state failure recovery
cost_governor.py	CostGovernor	Soft/hard budget caps
delegate_task.py	DelegationToken	HMAC-signed agent delegation
audit_log.py	AuditLog	Append-only governance log
agent_registry.py	AgentRegistry	Identity + trust tiers
schema_validator.py	SchemaValidator	Stdlib JSON Schema validation
compliance_mapper.py	ComplianceMapper	SOC2/GDPR evidence mapping
health_probe.py	HealthProbe	Composable health checks
output_validator.py	OutputValidator	Content safety filtering
capability_fence.py	CapabilityFence	Role-based capability boundaries
stride_mapper.py	StrideMapper	STRIDE threat analysis
lifecycle.py	GovernanceLifecycle	NIST AI RMF orchestration
contract_net.py	ContractNetManager	Multi-agent task allocation
convergence_guard.py	ConvergenceGuard	Instrumental convergence detection
behavior_monitor.py	BehaviorMonitor	Behavioral drift detection
lamport_clock.py	LamportClock	Distributed causal ordering
reasoning.py	ReasoningEngine	Structured governance reasoning
eal.py	EAL	Execution assurance
physical_governor.py	KinematicGovernor	Physical-AI safety limits
errors.py	HummblError	Typed error taxonomy
failure_modes.py	FailureMode	Failure classification
evolution_lineage.py	EvolutionLineage	eAI variant lineage tracking
agent_runner.py	--	End-to-end agent with governance stack
failure_injection.py	--	Chaos testing with kill switch + circuit breaker

Run them all:

for f in examples/*.py; do echo "=== $f ==="; python "$f"; done

OWASP Top 10 for Agentic Applications (2026) Engineering Mapping

The table below maps hummbl-governance primitives to the OWASP Top 10 for Agentic Applications. This is an engineering evidence map, not a third-party certification or attestation. Most rows below link to a primitive and its test suite; ASI04 is a supply-chain posture row for the package runtime surface.

OWASP Risk	Primitive(s)	Tests	How
ASI01 Agent Goal Hijack	KillSwitch	27	4-mode graduated shutdown (DISENGAGED → EMERGENCY). Survives process restart. Stops hijacked agents mid-execution.
ASI02 Tool Misuse	CapabilityFence	27	Allowlist/blocklist enforcement per tool. Agents cannot invoke tools outside their granted capabilities.
ASI03 Identity & Privilege Abuse	DelegationTokenManager, AgentRegistry	16 + 26	HMAC-signed scoped tokens with chain-depth limits. Identity registry with trust tiers and alias canonicalization.
ASI04 Supply Chain	Zero third-party runtime dependencies	—	Stdlib-only runtime surface reduces transitive dependency exposure. Package, build, publishing, CI, and optional/test tooling still require normal supply-chain review.
ASI05 Unexpected Code Execution	OutputValidator, InjectionDetector	49	Prompt injection detection, blocked-term filtering, and content validation before agent output reaches downstream systems.
ASI06 Memory & Context Poisoning	BusWriter, AuditLog	63 + 34	Append-only governance bus with content hashing. Tamper-evident audit log. Poisoned entries are detectable.
ASI07 Insecure Inter-Agent Comms	LamportClock, ContractNetManager	20 + 19	Hardened logical clocks for causal ordering. Contract Net protocol for structured multi-agent task allocation with bid verification.
ASI08 Cascading Failures	CircuitBreaker, HealthProbe	17 + 30	CLOSED/HALF_OPEN/OPEN state machine isolates failing components. Health probes detect degradation before cascade.
ASI09 Human-Agent Trust Exploitation	ReasoningEngine, ComplianceMapper	7 + 112	Structured decision traces explain why a governance decision was made. Compliance mapping to NIST/ISO provides external validation anchor.
ASI10 Rogue Agents	BehaviorMonitor, GovernanceLifecycle	20 + 17	Jensen-Shannon divergence detects behavioral drift from baseline. Lifecycle FSM enforces PROVISIONED → ACTIVE → SUSPENDED → DECOMMISSIONED transitions.

Current posture: 2081 collected tests, 34 implemented primitives, 7 MCP server entry points, and zero third-party runtime dependencies. Test-count and coverage claims require current receipts; see docs/public-claims.md.

For the formal governance primitive underlying all 10 mitigations, see The Governance Tuple (Bowlby, 2026).

Research

The evidence base behind hummbl-governance is documented in the AI Slop Crisis research corpus:

• Why Libraries, Not Platforms -- the architectural thesis behind stdlib-only, independently importable governance primitives
• Vendor Comparison Table -- how hummbl-governance compares to platform-locked alternatives across dependency count, modularity, and compliance coverage

Newsletter

Subscribe to the HUMMBL Slop Tracker for monthly AI governance intelligence: hummbl.substack.com

Read Issue #1 for the inaugural edition.

FAQ

How do I add a kill switch to my AI agent system?

Install hummbl-governance and use the KillSwitch class. It provides 4 graduated modes: DISENGAGED (normal operation), HALT_NONCRITICAL (stop non-essential tasks), HALT_ALL (stop everything except monitoring), and EMERGENCY (immediate full shutdown). Call ks.check_task_allowed("task_name") before each agent action.

from hummbl_governance import KillSwitch, KillSwitchMode
ks = KillSwitch()
ks.engage(KillSwitchMode.HALT_NONCRITICAL, reason="High error rate", triggered_by="monitor")

How do I track AI API costs and enforce budget limits?

Use CostGovernor with soft and hard caps. Record each API call with record_usage(), then call check_budget_status() to get an ALLOW, WARN, or DENY decision. The soft cap triggers warnings; the hard cap blocks further spending.

from hummbl_governance import CostGovernor
gov = CostGovernor(":memory:", soft_cap=50.0, hard_cap=100.0)
gov.record_usage(provider="openai", model="gpt-4", tokens_in=500, tokens_out=200, cost=0.02)

How do I implement delegation tokens for multi-agent AI systems?

Use DelegationTokenManager to create HMAC-SHA256 signed tokens that grant specific operations to specific agents. Tokens are scoped by issuer, subject, allowed operations, and an optional binding to a task and contract. Validate tokens before executing delegated actions.

from hummbl_governance import DelegationTokenManager
from hummbl_governance.delegation import TokenBinding
mgr = DelegationTokenManager(secret=b"shared-secret")
token = mgr.create_token(issuer="orchestrator", subject="worker", ops_allowed=["read", "write"],
                         binding=TokenBinding("task-1", "contract-1"))
valid, error = mgr.validate_token(token)

Does hummbl-governance work without any third-party packages?

Yes. Every runtime module uses only Python stdlib (3.11+). There are zero entries in the dependencies list in pyproject.toml. Test dependencies are isolated under project.optional-dependencies.test in pyproject.toml. This means no runtime dependency conflicts, reduced transitive dependency risk, and fast installs.

How do I generate compliance evidence for SOC2 or GDPR from my AI system?

Use ComplianceMapper to map governance audit log entries to compliance framework controls. Pass your AuditLog entries through the mapper to produce a ComplianceReport that links each governance event to the relevant SOC2, GDPR, or OWASP control. Use StrideMapper for threat-level analysis of agent interactions.

from hummbl_governance import ComplianceMapper, AuditLog
mapper = ComplianceMapper()
report = mapper.map_events(audit_entries)  # Returns ComplianceReport with control mappings

Multi-Agent Example

Delegation tokens and agent identity for multi-agent systems:

from hummbl_governance import DelegationTokenManager, AgentRegistry
from hummbl_governance.delegation import TokenBinding

# Agent registry: identity + trust tiers
registry = AgentRegistry()
registry.register_agent("orchestrator", trust="high")
registry.register_agent("worker-1", trust="medium")
registry.add_alias("orch-1", "orchestrator")
registry.canonicalize("orch-1")  # -> "orchestrator"

# Delegation tokens: HMAC-signed, scoped, chain-depth-limited
mgr = DelegationTokenManager(secret=b"shared-secret")
token = mgr.create_token(
    issuer="orchestrator", subject="worker-1",
    ops_allowed=["read", "write"],
    binding=TokenBinding("task-1", "contract-1"),
)
valid, error = mgr.validate_token(token)  # (True, None)

# Worker can now prove it's authorized for "read"/"write" on "task-1"
# Token is tamper-evident — any modification invalidates the signature

MCP Servers

hummbl-governance ships three Model Context Protocol servers that expose governance primitives as tools to any MCP-compatible client (Claude Code, Claude Desktop, etc.).

hummbl-governance (core)

{
  "mcpServers": {
    "hummbl-governance": {
      "command": "python",
      "args": ["/path/to/hummbl-governance/mcp_server.py"],
      "env": {
        "GOVERNANCE_STATE_DIR": "/path/to/state"
      }
    }
  }
}

10 tools: governance_status, kill_switch_status, kill_switch_engage, kill_switch_disengage, circuit_breaker_status, cost_budget_check, cost_record_usage, audit_query, compliance_report, health_check

hummbl-compliance

{
  "mcpServers": {
    "hummbl-compliance": {
      "command": "python",
      "args": ["/path/to/hummbl-governance/mcp_compliance.py"],
      "env": {
        "GOVERNANCE_AUDIT_DIR": "/path/to/audit"
      }
    }
  }
}

5 tools: nist_map_controls, soc2_assess, iso_crosswalk, stride_analysis, compliance_evidence_export

hummbl-sandbox

{
  "mcpServers": {
    "hummbl-sandbox": {
      "command": "python",
      "args": ["/path/to/hummbl-governance/mcp_sandbox.py"],
      "env": {
        "SANDBOX_STATE_DIR": "/path/to/sandbox"
      }
    }
  }
}

This server is part of hummbl-governance and is launched through the hummbl-sandbox-mcp entry point. It is not related to the third-party npm package named agent-sandbox.

5 tools: sandbox_create, sandbox_check, sandbox_validate_output, sandbox_status, sandbox_destroy

All three servers use stdio JSON-RPC and have zero third-party dependencies.

Design Principles

• Zero third-party runtime dependencies -- stdlib only (Python 3.11+)
• Thread-safe -- all modules use appropriate locking
• Configurable -- no hardcoded agent names or paths
• Independently importable -- each module works standalone

Development

python -m venv .venv && source .venv/bin/activate
pip install -e ".[test]"
python -m pytest tests/ -v

HUMMBL Ecosystem

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

Repo	Purpose
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
hummbl-agent	Governed control plane for AI agent systems
hummbl-bibliography	Bibliography for the HUMMBL cognitive framework
agentic-eng-patterns	Stdlib-only safety patterns for agentic AI systems

Links

• PyPI
• GitHub
• hummbl.io
• Linktree
• HUMMBL Slop Tracker (Substack)

License

Apache 2.0. Copyright 2026 HUMMBL, LLC.

---

Author

Reuben Bowlby — Founder, HUMMBL

• GitHub: @hummbl-dev
• X/Twitter: @ReubenBowlby
• LinkedIn: reuben-bowlby
• Website: hummbl.io