Agent Harness Tutorial · End to end · Zoom level 2
Check point 8: Before the reply
The very last look. It runs on direct answers too, so even approved templates are checked.
Where: on every path, templates included.
| Check | Catches | Kind | Provided by | If it fails |
|---|---|---|---|---|
| no_promise_of_cover | a promise of cover | rule | agent | safe reply |
| no_pii_echo | personal data | rule | agent | safe reply |
| money_in_rupees | the wrong currency | rule | agent | safe reply |
| name the reference | a new claim without its CLM- number | rule | agent | number added |
Only the three rules that need nothing but the reply text run here. The grounded rules need tool results, so they run after the model only.
How it is done: the real code
Cut from the live repositories each time the site is built, so it always matches what runs.
agent_harness/entrypoint/ending.py:127screened(): the last look, on every routedef screened(
result: TurnResult,
identity: Identity,
rules: Mapping[pol.Position, tuple[pol.Rule, ...]] | None = None,
) -> TurnResult:
"""Every reply passes the reply guardrails before the customer reads it —
whichever route produced it (F-020). One point, so the next template edit on
any route is screened without anyone remembering to screen it. The verdict
and the rule that fired are on the `agent.policy` span `enforce` opens.
The agent's **configured** rules, not only the built-in ones: this read the
defaults whatever it was given, so a deployment that added a reply rule was
screened by the rules it had not configured (F-027)."""
text = result.customer_message if isinstance(result, Failed) else result.reply
verdict = pol.enforce(
pol.Context(position=pol.Position.REPLY, identity=identity, text=text),
None if rules is None else rules.get(pol.Position.REPLY),
)
if not verdict.blocked:
return result
# The words are replaced; what the turn *did* is not. An escalation that was
# raised stays raised and an approval stays pending — replacing the result
# type would drop the handoff the reply was about.
if isinstance(result, Failed):
return result.model_copy(update={"customer_message": pol.SAFE_REPLY})
if isinstance(result, Completed):
# A completion carries nothing but its words, and the words were
# refused — so the result is a refusal. It used to stay `Completed`
# with a `REFUSED` termination, which a caller branching on the type
# read as a success (F-026, AHC-0017).
return Refused(reply=pol.SAFE_REPLY, reason=verdict.reason, rule_id=verdict.rule)
return result.model_copy(update={"reply": pol.SAFE_REPLY})