agent-execution-guard

agent-execution-guard

Deterministic ALLOW / HOLD / DENY execution boundary for AI agent actions.

pip install agent-execution-guard

---

The problem

AI agents execute actions without structural constraints. Prompts can be bypassed. Guardrails can be reasoned around.

This library puts a deterministic gate between decision and execution.

---

Quickstart

from datetime import datetime, timezone
from agent_execution_guard import ExecutionGuard, Intent, ALLOW_ALL

guard = ExecutionGuard()

intent = Intent(
    actor="agent.finance",
    action="wire_transfer",
    payload="wire_transfer amount=50000 to=external",
    timestamp=datetime.now(timezone.utc),
)

# policy is required — no policy = no execution (fail-closed)
# use ALLOW_ALL to skip identity check and rely on risk scoring only
try:
    result = guard.evaluate(intent, policy=ALLOW_ALL)
    print(result.risk_score)  # 0–100

except Exception as e:
    print(e)  # DENY or HOLD

---

Example 1 — DENY

from agent_execution_guard import ExecutionGuard, Intent, GuardDeniedError, ALLOW_ALL
from datetime import datetime, timezone

# halt_threshold=39: wire_transfer scores 40 → DENY
guard = ExecutionGuard(halt_threshold=39)

intent = Intent(
    actor="agent.finance",
    action="wire_transfer",
    payload="wire_transfer amount=50000 to=external",
    timestamp=datetime.now(timezone.utc),
)

try:
    guard.evaluate(intent, policy=ALLOW_ALL)

except GuardDeniedError as e:
    print(f"Denied: {e.reason}")
    print(f"Proof:  {e.boundary_id}")

Output:

Denied: General command
Proof:  3f9a1c2d-...

Every denial issues a signed proof. The agent cannot retry past it.

---

Example 2 — Severity-adaptive gate

from agent_execution_guard import ExecutionGuard, Intent, SystemSeverity, GuardDeniedError
from datetime import datetime, timezone

guard = ExecutionGuard()

intent = Intent(
    actor="campaign_ai",
    action="aggressive_targeting",
    payload="launch targeted ad campaign segment=undecided_voters",
    timestamp=datetime.now(timezone.utc),
)

# score=0.60 → COOLDOWN state → threshold drops 80 → 30 → risk(40) > 30 → DENY
high = SystemSeverity(score=0.60, source="risk_model")
low  = SystemSeverity(score=0.10, source="baseline")

try:
    guard.evaluate(intent, severity=high, policy=ALLOW_ALL)
except GuardDeniedError:
    print("DENY  (COOLDOWN — threshold=30)")

result = guard.evaluate(intent, severity=low, policy=ALLOW_ALL)
print("ALLOW (ACTIVE  — threshold=80)")

Output:

DENY  (COOLDOWN — threshold=30)
ALLOW (ACTIVE  — threshold=80)

Same action. Different severity state. Different outcome.

---

Policy guard

policy is required. No policy = no execution (fail-closed by default).

# policy.yaml
defaults:
  unknown_agent:  DENY
  unknown_action: DENY

identity:
  agents:
    - agent_id: "agent.finance"
      allowed_actions:
        - action: "wire_transfer"

import yaml
with open("policy.yaml") as f:
    policy = yaml.safe_load(f)

guard.evaluate(intent, policy=policy)
# unknown agent → immediate DENY, signed proof issued

To skip identity check and rely on risk scoring only:

from agent_execution_guard import ALLOW_ALL

guard.evaluate(intent, policy=ALLOW_ALL)

---

How it works

Intent → Risk Score → Severity State → Guard Decision → Signed Proof

States

Outcomes

---

Cryptographic proof

Every decision — ALLOW and DENY — is signed and ledgered.

• ED25519 signatures
• Append-only NDJSON hash chain
• Offline verification, no external dependencies

---

Requirements

• Python 3.10+
• cryptography>=42.0.0

Optional:

pip install pyyaml             # policy.yaml support
pip install opentelemetry-api  # OTel span export

---

Roadmap

• Risk scoring + ALLOW / DENY
• ED25519 cryptographic proof + hash-chain ledger
• Severity-driven state machine (ACTIVE / OBSERVE / COOLDOWN)
• Policy guard (unknown agent/action → DENY)
• HOLD state + human approval checkpoint
• LangChain adapter (GuardedTool)
• OTel-native decision trail export
• MCP integration

---

License

Apache 2.0