agec 0.1.0

Authorized Governance Execution Context for AI agents

AGEC

Pre-execution governance layer for AI agents.

Every AI agent action should be authorized before execution. AGEC validates intent, semantic context, execution path, and data processing permissions before any tool call is executed. It provides deterministic authorization, replayable decisions, and auditability for autonomous AI agents.

---

Installation

pip install agec

Quick Start

from agec import guard, AGECBlockedError

@guard(
    intent="send_email",
    purpose="customer_support",
    allowed_tools=["gmail.send"],
    legal_basis="consent",
)
def send_email() -> str:
    return "Email sent."


@guard(
    intent="transfer_money",
    purpose="unknown",
    allowed_tools=["bank.transfer"],
    intent_confidence=0.48,         # below the 0.7 threshold
)
def unsafe_transfer() -> str:
    return "Transferred $1M."


print(send_email())                 # → Email sent.

try:
    unsafe_transfer()
except AGECBlockedError as exc:
    print("AGEC BLOCKED EXECUTION")
    print(f"Reason:   {exc.reason}")
    print(f"Audit ID: {exc.agec.agec_id}")

If validation fails, execution is blocked before the wrapped function runs.

---

Why AGEC?

Traditional authorization systems answer:

Can this identity access this resource?

AGEC answers a different question:

Should this exact action execute right now?

AGEC introduces a mandatory governance layer between agent planning and tool execution — combining intent validation, policy enforcement, and tamper-evident audit logging in a single decorator.

---

What AGEC Validates

Check	What it enforces
Intent	Declared intent must be in the policy allowlist
Confidence	Intent confidence must meet the minimum threshold
Execution path	Every tool step must be explicitly permitted
Purpose	Data processing purpose must be allowed
Legal basis	GDPR-aligned legal basis must be declared and allowed
Data categories	Blocked data categories are rejected before execution
Expiry (TTL)	Stale contexts are cancelled automatically

---

Architecture

User
  │
AI Agent (planning)
  │
AGEC  ◄─── Policy + Validator
  │              │
  │         AuditLog ──► audit.jsonl (optional)
  │
Tool Execution

---

Lower-Level API

from agec import AGEC, Intent, ExecutionPath, DataPermissions, Policy, AGECValidator

policy = Policy(
    allowed_intents=["send_email"],
    allowed_tools=["gmail.send"],
    allowed_purposes=["customer_support"],
    allowed_legal_bases=["consent", "contract"],
)

agec = AGEC(
    intent=Intent(type="send_email", confidence=0.95),
    context={"user_id": "123"},
    execution_path=ExecutionPath(path_id="email_path", steps=["gmail.send"]),
    data_permissions=DataPermissions(
        purpose="customer_support",
        legal_basis="consent",
        allowed_operations=["send"],
        data_categories=["email"],
    ),
)

validator = AGECValidator(policy)
result = validator.validate(agec)

print(result.allowed)   # True
print(result.reason)    # AGEC validation passed.
print(agec.status)      # AGECStatus.ACTIVE

Persisting the Audit Log

from agec import AuditLog

log = AuditLog()
# ... pass log to AGECValidator(policy, audit_log=log) ...

# Save all recorded events to disk
log.save_json("audit.jsonl")

# Reload later for replay or compliance review
restored = AuditLog.load_json("audit.jsonl")

---

Roadmap

• OpenAI Agents SDK adapter
• LangGraph adapter
• CrewAI adapter
• AutoGen adapter
• CLI demo runner
• Policy manifest (YAML/JSON) support
• Replayable audit log (JSON persistence)
• Deterministic execution path hashing

---

Contributing

See CONTRIBUTING.md.

Changelog

See CHANGELOG.md.

License

Apache-2.0