Agent Runtime Governance

The action a reviewer approved is the action that executes. A side effect with an uncertain outcome is recorded as UNKNOWN, and automatic reuse stays blocked while that idempotency record is retained.

Two failures motivate this project. An agent retries a payment call after a network timeout and charges a customer twice. A reviewer approves one tool call, and something different executes because arguments or governance state changed between approval and execution. Agent Runtime Governance is an embeddable, framework-agnostic Python runtime that closes both gaps at the final execution boundary - the moment a proposed tool call is about to touch the real world within the configured storage guarantees. It runs inside the agent stack you already use rather than replacing it. The default in-memory idempotency store has bounded TTL/LRU retention and does not survive restarts; longer-lived protection requires a durable store. Deterministic UNKNOWN reconciliation is planned for v0.7.

Every shipped guarantee links to its regression test:

Shipped in v0.5.1	Evidence
Approval is bound to request ID, tool name, argument digest, risk tier, policy version/digest, subject, tenant, identity issuer, and request expiry. A mismatch or expired request/decision is rejected before the tool body runs	test_decision_binding_rejects_wrong_request_or_arguments
Caller-supplied metadata cannot forge approval, policy, or identity state	test_caller_metadata_cannot_forge_required_approval
A mutating call with an uncertain outcome is recorded as UNKNOWN instead of being retried	test_cancellation_propagates_and_marks_context_unknown
Idempotency keys are mandatory for idempotent mutating tools and bound to the parameter fingerprint	test_idempotency_key_is_bound_to_parameter_fingerprint
JSONL audit is hash-chained and signable; deletion and tampering are detected	test_jsonl_audit_hash_chain_detects_deletion_and_tamper

v0.6 binds one immutable action across the final execution boundary. Every claim below is covered by a regression test or recorded benchmark:

Shipped in v0.6.0	Evidence
The exact frozen parameter snapshot covered by action_digest reaches the tool body and cannot be replaced by middleware or hooks	test_exact_bound_snapshot_reaches_tool_and_audit, test_middleware_cannot_replace_or_mutate_bound_action
Key, policy, and external-precondition drift fail before tool entry	test_key_rotation_fails_before_tool_entry, test_policy_identity_mismatch_fails_closed, test_precondition_change_fails_before_tool_entry
Approval and idempotency consume action_digest; v0.5 records remain readable but cannot authorize or satisfy a contracted v0.6 action	test_approval_is_bound_to_action_digest, test_v05_approval_fails_closed_for_contracted_tool, test_v05_compatibility.py
Success and terminal exception/timeout/cancellation paths retain the same action identity	test_exact_bound_snapshot_reaches_tool_and_audit, test_exception_and_timeout_keep_bound_action_in_audit, test_cancellation_keeps_bound_action_in_audit
Audit evidence carries the action identity without duplicating raw parameters, and OpenTelemetry exports the same contract/action attributes	test_audit_bound_action_never_duplicates_raw_parameters, test_opentelemetry_exports_bound_action_identity
At 1,000 requests and concurrency 100 on the recorded Windows/Python 3.12 host, the median of three alternating paired runs measured action bind plus verification at 1.553x mean, 1.720x p95, 1.850x p99, and 1.040x peak traced memory versus its strict baseline	v0.6.0-windows-python312.json, v0.6.0 budget

The v0.7-v1.0 direction adds deterministic reconciliation for UNKNOWN outcomes and portable evidence. The staged plan and its exit criteria are in docs/production-roadmap.md; planned features are never presented as shipped.

Quick start

pip install agent-runtime-governance

from pathlib import Path

from agent_runtime_governance import ExecutionMode, Runtime

runtime = Runtime()
WORKSPACE_ROOT = Path.cwd().resolve()

@runtime.tool(execution_mode=ExecutionMode.READ_ONLY)
def read_file(path: str) -> str:
    candidate = (WORKSPACE_ROOT / path).resolve()
    try:
        candidate.relative_to(WORKSPACE_ROOT)
    except ValueError as exc:
        raise ValueError("path must remain inside the workspace") from exc
    return candidate.read_text(encoding="utf-8")

print(read_file("README.md"))

Add deterministic rules, semantic review, human decisions, and signed audit:

import os

from agent_runtime_governance import (
    ApprovalMiddleware, AuditMiddleware, HumanDecisionProvider,
    InvocationOptions, JSONLAuditSink, LLMMiddleware, RiskTier,
    Rule, RuleMiddleware, Runtime,
)

runtime = Runtime([
    RuleMiddleware([Rule("explicit-wipe", r"\bwipe\s+all\b", "bulk wipe is forbidden")]),
    LLMMiddleware(lambda ctx: True),
    ApprovalMiddleware(HumanDecisionProvider(lambda ctx, request: True)),
    AuditMiddleware(
        JSONLAuditSink("audit.log", sign_key=os.environ["ARG_AUDIT_HMAC_KEY"])
    ),
])

@runtime.tool(risk=RiskTier.HIGH, requires_approval=True)
def delete_file(path: str) -> bool:
    return True

delete_file(
    "old.log",
    _governance=InvocationOptions(input_text="remove the old application log"),
)

How it compares

Human approval, policy interception, audit, retries, and telemetry are industry baseline; this project does not claim to have invented them. The narrow claim is the governed execution boundary. Rows cite each project's own public documentation and describe scope, not quality.

Alternative	What it provides	Where this project is narrower and deeper
Framework-native approvals: OpenAI Agents SDK, LangGraph, pydantic-ai deferred tools	Pause-and-approve flows inside one framework	One framework-neutral runtime where approval is revalidated against the normalized call immediately before execution, combined with idempotent commit and UNKNOWN semantics
Microsoft Agent Governance Toolkit	Broad policy, identity, sandboxing, and SRE toolkit	Its versioned limitations document records that audit captures attempts and allow/deny results, with outcome attestation planned; this project's entire scope is that commit-and-outcome boundary
Durable execution engines: Temporal, Restate, DBOS	Durable workflow state with automatic retries	The governance-layer default is different: an uncertain side effect is recorded as UNKNOWN and is not automatically retried while its idempotency record is retained; deterministic reconciliation is planned for v0.7, and human approval binds to the action identity
Content guardrails: NeMo Guardrails, Guardrails AI	Input, output, and dialog rails	A different layer that composes with this one; this project governs the side-effect boundary, not prompts or content

Prompt instructions are useful guidance, but they are not an authorization boundary. This project places deterministic policy, explicit decisions, and audit outside the model while remaining independent from agent planning and model providers.

Architecture

Tool Registry
     |
     v
Immutable ExecutionContext
     |
     v
Runtime Pipeline
  Prepare -> Bind Action -> Policy -> Human Decision
     |
     v
Action Digest -> Idempotency -> Executor Revalidation -> Tool
     |
     v
Terminal Context -> Evidence-safe Audit Snapshot

Gating middleware may stop execution and fails closed. Observing middleware cannot grant permission and normally cannot interrupt an allowed tool. An audit sink explicitly configured as critical is the exception: failed delivery stops the call because an unaudited privileged action is not allowed. Earlier denials cannot be overridden by later middleware.

Design principles

1. Deterministic first - code and policy establish the execution boundary.
2. Policy over prompt - prompts guide behavior; policies authorize actions.
3. Defense in depth - independent controls can only tighten a decision.
4. Framework agnostic - the runtime does not own planning or model calls.
5. Human in the loop - applications supply the decision channel.
6. Observability by default - each transition produces traceable state.
7. Immutable context - middleware returns a new context instead of mutating shared state.

Examples

• examples/standalone_demo.py: framework-free pipeline
• examples/cli_approval_demo.py: interactive human decision provider
• examples/langgraph_integration.py: LangGraph tool-node integration
• examples/openai_agents_integration.py: OpenAI Agents SDK tool integration
• examples/crewai_integration.py: CrewAI decorated tool
• examples/agno_integration.py: Agno function tool
• examples/llamaindex_integration.py: LlamaIndex FunctionTool
• examples/autogen_integration.py: Microsoft AutoGen FunctionTool
• examples/strict_action_contract.py: sealed contracted side-effecting runtime
• scripts/replay.py: print the snapshots for a trace from JSONL audit data

Each framework adapter is intentionally an example rather than a runtime dependency. The governed function remains ordinary Python and can be wrapped by the framework's native tool interface.

Engineering controls

v0.2 adds immutable pipeline composition, lifecycle hooks, Python-native policy, metrics, retries, timeouts, and optional OpenTelemetry export:

from agent_runtime_governance import (
    MetricsMiddleware, Pipeline, PolicyMiddleware, RetryMiddleware,
    SimplePolicy, TimeoutMiddleware,
)

pipeline = Pipeline([
    PolicyMiddleware(SimplePolicy(admin_only={"restart_service"})),
    RetryMiddleware(max_attempts=2),
    TimeoutMiddleware(5.0),
    MetricsMiddleware(),
])
runtime = Runtime(pipeline)

@runtime.before_tool
def attach_region(ctx):
    return ctx.evolve(metadata={**ctx.metadata, "region": "cn-beijing"})

Pipeline edits return a new Pipeline; a live runtime is never mutated behind concurrent calls. Hooks may enrich context but cannot change status or decisions.

Policies, snapshots, and regression

Load a versioned policy with separate semantic and exact-artifact digests:

from agent_runtime_governance import Runtime, YAMLPolicyLoader

document = YAMLPolicyLoader.load("examples/policy.yaml")
runtime = Runtime([document.artifact_middleware()])
print(document.version, document.digest, document.artifact_digest)

Duplicate tool entries, unknown fields, invalid risks, and unsafe YAML tags are rejected. YAML remains a configuration format over the deliberately small SimplePolicy model; it is not a general policy language. digest identifies normalized policy semantics for compatibility and drift comparison; artifact_digest identifies the exact loaded bytes and is the value strict production profiles must bind.

SnapshotMiddleware records immutable lifecycle snapshots. ReplayDebugger prints timelines and field-level diffs, while EvaluationSuite runs governance without executing tools. PolicyDriftDetector reapplies deterministic policy to the same recorded request identity and reports decision or risk changes. Runtime.areplay() is deliberately non-authoritative: historical identity fields are analysis inputs, verification metadata is stripped, no tool runs, and no BoundAction is created. Use apreview() with current trusted identity claims when a fresh bound action preview is required.

python scripts/trace_debug.py snapshots.jsonl TRACE_ID
python scripts/trace_debug.py snapshots.jsonl TRACE_ID --diff 0 1
python scripts/trace_debug.py snapshots.jsonl TRACE_ID --mermaid

Plugins and integrations

Plugins register components during construction; the built runtime remains immutable:

from agent_runtime_governance import OPAClient, OPAPlugin, PluginManager

manager = PluginManager()
manager.load(OPAPlugin(OPAClient("http://localhost:8181", "agents/tools/allow")))
runtime = manager.build()

Third-party plugins can publish the Python entry-point group agent_runtime_governance.plugins. Entry points execute Python code and must be treated as trusted dependencies. The project does not download plugins or provide a marketplace.

Integration	Extra	Behavior
Prometheus	prometheus	Terminal status and duration; no trace/user labels
Slack	none	Denial/failure notifications; official HTTPS webhooks only
OPA	none	Minimal decision input; fail closed by default
LangGraph	langgraph	Governed function in a graph node
OpenAI Agents SDK	openai-agents	Governed async function tool
CrewAI	crewai	Decorated governed async tool
Agno	agno	Typed governed Python function
LlamaIndex	llamaindex	Governed FunctionTool async function
Microsoft AutoGen	autogen	Governed FunctionTool async function

Install only the integrations an application uses:

pip install "agent-runtime-governance[yaml,prometheus,crewai]"

Production reliability

v0.5 establishes the production reliability baseline:

• idempotent tool execution with mandatory keys and in-memory or SQLite stores;
• durable approval recovery with leased reserve/commit/release transitions;
• absolute deadline propagation through admission, middleware, identity, tools, and idempotency waits;
• cancellation handling that records mutating in-flight work as UNKNOWN;
• trusted HMAC identity envelopes with replay protection;
• contract validation for parameters, results, and payload size limits;
• reliable JSONL and SQLite audit sinks with hash-chain verification;
• bounded concurrency, external integration circuit breakers, and fault tests.

v0.6 adds strict startup inventory and sealing, one BoundAction.action_digest across approval/idempotency/execution/audit, executor-boundary revalidation, policy and precondition identity checks, and versioned v0.5 migration readers. See docs/migration-v0.6.md and the runnable strict_action_contract.py example.

The example hashes the exact policy artifact in examples/policies/, passes that identity through PolicyMiddleware and ProductionProfile, seals the runtime, and persists signed audit plus idempotency state:

python -m pip install -e .
$env:ARG_IDENTITY_DIGEST_KEY = python -c "import secrets; print(secrets.token_urlsafe(32))"
$env:ARG_AUDIT_HMAC_KEY = python -c "import secrets; print(secrets.token_urlsafe(32))"
$env:ARG_STATE_DIR = ".arg-state-example"
python examples/strict_action_contract.py
Remove-Item -Recurse -Force $env:ARG_STATE_DIR

python -m pip install -e .
export ARG_IDENTITY_DIGEST_KEY="$(python -c 'import secrets; print(secrets.token_urlsafe(32))')"
export ARG_AUDIT_HMAC_KEY="$(python -c 'import secrets; print(secrets.token_urlsafe(32))')"
export ARG_STATE_DIR=.arg-state-example
python examples/strict_action_contract.py
rm -rf -- "$ARG_STATE_DIR"

StaticIdentityProvider in this example represents one identity already authenticated by a trusted single-service host boundary. It must not be used to authenticate arbitrary multi-user requests; use a validating identity provider at that boundary in production.

The production smoke suite starts real Docker services for OPA and the OpenTelemetry Collector, exports through OTLP HTTP, scrapes a real HTTP /metrics exposition endpoint, and can run a local Kind smoke with a pinned node image:

python integration/production_smoke.py --skip-kind
python integration/production_smoke.py

Production deployments must configure a trusted identity provider with require_verified_identity=True; caller-supplied user, tenant, and permissions fields are compatibility inputs, not a trust boundary. SQLite stores coordinate processes on one host and require a distributed adapter for multi-host deployments.

Release verification

Each release is verified before publishing: the full test matrix on Python 3.10-3.13, Docker-backed integration smoke, an isolated dependency audit, and wheel/sdist installation from clean environments, with the SPDX SBOM, SHA256 checksums, and GitHub provenance attached to the release. The complete point-in-time record for v0.5.1 is in docs/release-verification.md.

Benchmarks measure incremental runtime overhead for baseline, Rule, OPA, Audit, OpenTelemetry, 10-middleware pipelines, and paired strict runtimes with and without action binding:

python benchmarks/benchmark_runtime.py --requests 100,500,1000 --concurrency 100

The committed Windows/Python 3.12 measurement is available in benchmarks/results/v0.5.0-windows-python312.json. Its 100-waiter, zero-hold FIFO admission test measured 2.93 ms p99 wait time. This point-in-time result is regression evidence for that machine, not a cross-platform latency SLA.

The v0.6 final measurement, earlier pre-release measurement, and enforced paired budget are in benchmarks/results/v0.6.0-windows-python312.json, benchmarks/results/v0.6.0-rc-windows-python312.json, and benchmarks/budgets/v0.6.0.json.

Releases

Version	Scope
v0.1.0	Immutable context, registry, rule/LLM/approval/audit middleware, basic replay
v0.2.0	Hooks, Python policy, metrics, retry, timeout, and OpenTelemetry bridge
v0.3.0	Strict YAML policy, snapshots, replay diff, evaluation, and policy drift
v0.4.0	Trusted plugins, Prometheus, Slack, OPA, and six framework integrations
v0.4.1	Mandatory linked issues and merge-policy verification
v0.4.2	Fail-closed CodeRabbit review verification for the current commit
v0.5.0	Production reliability: idempotency, identity, durable approvals, audit, deadlines, cancellation, contracts, real integration smoke
v0.5.1	Security hardening: caller metadata isolation and exact approval binding
v0.6.0	Immutable action contracts from admission through approval, idempotency, executor revalidation, telemetry, and audit

Released versions are preserved as immutable Git tags. See CHANGELOG.md for the detailed compatibility and security notes.

Non-goals

• Agent planning, prompting, model routing, or memory
• Multi-agent communication or distributed execution
• Approval UI or a hosted control plane
• A general policy language, plugin marketplace, or mutable live pipeline
• Production-grade time-travel debugging
• Arbitrary condition evaluation, policy inheritance, or conflict resolution

See ARCHITECTURE.md for invariants and ROADMAP.md for the deliberately staged scope. Security reports and integration boundaries are documented in SECURITY.md. Contributor workflow and plugin contribution rules are documented in CONTRIBUTING.md. Production configuration and recovery are covered in docs/production.md, and the release procedure is defined in RELEASING.md. The evidence-backed v0.6-v1.0 strategy is specified in docs/production-roadmap.md.

Development

python -m pip install -e ".[dev]"
pytest
python integration/production_smoke.py --skip-kind
python -m build

Python 3.10+ is supported. The core package depends on filelock and jsonschema; optional integrations are installed via extras.

Pull requests must link an existing issue in this repository with a closing keyword such as Fixes #123. A repository workflow closes unlinked pull requests automatically. Pull requests to main must also pass every CI job and the CodeRabbit status, receive a verified CodeRabbit approval for the current head commit, and resolve blocking reviews. A skipped, rate-limited, stale, or missing CodeRabbit review fails closed. GitHub does not allow a pull request author to approve their own change, so the single-maintainer phase has no human approval count; one code-owner approval and last-push approval become mandatory when a second maintainer is added. Maintainers still use the same issue and pull request flow as external contributors, and administrators cannot push directly to main or bypass protection.

License

MIT