Account setup
Four steps to a governed agent.
- 01
Register at
ceadly.mewith a work email. - 02
Receive an API key (format:
ceadly_live_sk_...) — shown once, store in a secrets manager. - 03
The first registrant is automatically assigned the CISO role.
- 04
No policy required to start: agents can create checkpoints immediately. Without a published policy, criticality is taken directly from each action's declared
@guard(criticality=...)value, with a MEDIUM floor (a declared LOW still requires human review). Publish a policy in-dashboard whenever you want full control over routing.
SDK installation
One dependency. One env var.
Install the Python SDK:
pip install ceadlySet your API key (defaults to https://api.ceadly.me; override with CEADLY_API_URL only for self-hosting):
export CEADLY_API_KEY=ceadly_live_sk_your_key_hereGuarding a function
One decorator. Full governance.
Decorate any function that performs a consequential action. The decorator blocks until a human approves, denies, or the timeout expires.
from ceadly import guard, CeadlyActionRejected, CeadlyActionTimedOut
@guard(
agent_id="billing-agent-001",
legal_responsible_person="[email protected]",
legal_responsible_title="VP Finance",
agent_owner="finance-team",
criticality="high",
timeout_seconds=600,
reversible=False,
)
def send_renewal_emails(customer_segment: str, count: int) -> dict:
# Runs ONLY after human approval.
...
return {"sent": count}Handle the two terminal outcomes explicitly:
try:
result = send_renewal_emails(customer_segment="lapsed_90_days", count=4217)
except CeadlyActionRejected as e:
# A human explicitly said no.
...
except CeadlyActionTimedOut as e:
# No decision arrived in time.
...Decorator parameter reference
Every parameter. No surprises.
| Parameter | Required | Description |
|---|---|---|
| agent_id | Required | Stable identifier for this agent (e.g., billing-agent-001). Used for routing, audit, and policy matching. |
| legal_responsible_person | Required | Email of the named human legally accountable for this agent's actions. Missing this suspends the agent entirely — it cannot create checkpoints until set in the dashboard. |
| legal_responsible_title | Required | Title of the legally responsible person (e.g., VP Finance, CISO). |
| agent_owner | Required | Team or org unit that owns the agent (e.g., finance-team, platform-infra). Used for routing and policy scope. |
| criticality | Required | One of LOW, MEDIUM, HIGH, REGULATORY. Drives reviewer SLA, escalation path, and whether auto-approve on timeout is permitted. |
| timeout_seconds | Optional | Maximum seconds to wait for a human decision. Defaults to policy default if omitted. If exceeded, raises CeadlyActionTimedOut. |
| reversible | Optional | Informational only. true means the action can be rolled back; false means it cannot. Does not change enforcement behavior. |
Criticality behavior
Criticality drives enforcement — not just policy.
TRUST DIFFERENTIATOR
HIGH and REGULATORY criticality levels cannot be configured to auto-approve on timeout. This is enforced in the SDK and API layer, not just a policy recommendation. No configuration drift, no accidental bypass — the gate holds.
| Criticality | Reviewer SLA | Escalation | Auto-approve on timeout | Notes |
|---|---|---|---|---|
| LOW | Best effort | None | Allowed | Low-risk actions. Reviewer notified but no SLA enforcement. |
| MEDIUM | Policy default (e.g., 4h) | To agent_owner after SLA | Allowed | Standard business actions. Escalation follows policy. |
| HIGH | Policy default (e.g., 1h) | To legal_responsible_person after SLA | Forbidden | High-consequence actions. Cannot be configured to auto-approve on timeout — enforced in code, not just policy. |
| REGULATORY | Policy default (e.g., 30m) | To legal_responsible_person + CISO after SLA | Forbidden | Regulated actions. Cannot be configured to auto-approve on timeout — enforced in code, not just policy. Audit trail is mandatory. |
Agent lifecycle
No accountable human = no action.
Agents start in NEEDS_SETUP if not pre-registered in the dashboard. They cannot create checkpoints untillegal_responsible_person and legal_responsible_titleare confirmed in the dashboard.
This is a feature, not a limitation: nobody legally accountable means the agent cannot act. This is not configurable.
Troubleshooting quick-reference
Common blockers, solved in seconds.
QDo I need a policy before agents can create checkpoints?
@guard(criticality=...) value — with a minimum MEDIUM enforced (a declared LOW is never allowed to skip human review). Configure a policy on the Policies page to take full control over routing instead of relying on the fallback.QAgent won't create checkpoints
legal_responsible_person and legal_responsible_title are set on the agent in the dashboard.QSDK can't connect
CEADLY_API_KEY is set and valid. For self-hosted, check CEADLY_API_URL.Limitations
Honest about what we don't do (yet).
Python-only SDK today. Other language bindings are on the roadmap.
No external blockchain-style anchoring. Integrity is provided by internal cryptographic hashing (SHA3-256) and append-only database enforcement. The Governance Evidence Package add-on adds daily public-ledger commitments for external verifiability.
