Gates

A Gate is a decision point in an Op: the results of one or more Guards determine what happens next. Guards check; Gates decide. The idea, and the word, are those of the Intelligence Hub whitepaper (section 5.2).

Not every failure is the same — some need a correction, some a review, some an escalation, some a stop — and a Gate writes that decision into the workflow. It is also where the autonomy an Op was granted ends: the point at which a person can interpret an output, decline it, override it or stop the work.

# gates/low-confidence-review.yaml
id: low-confidence-review
version: 0.0.1
name: Low Confidence Review

stage: post_run
guards:
  - confidence-guard:0.0.1
when: confidence < 0.80
then: human_review_required
otherwise: proceed
reviewers:
  - op-owner

Fields

Field	Meaning
id, version, name, description	As on every spec.
stage	The stage of an Op the decision is made at.
guards	The Guards whose results the Gate reads.
when	The condition, on the signals its Guards report. always for a Gate every run goes through. Required.
then	What happens when the condition holds. Required.
otherwise	What happens when it does not; proceed by default.
reviewers	The roles the decision is handed to. Required when either then or otherwise hands it to a person.
max_retries	How many times retry may be decided. Required, at least 1, when it can be.
enabled	Whether an Op may name it today.

Conditions

A condition is a signal on its own — true when the Guard reported it — or a comparison of one with a number, true, false or a word; several are joined by and or or.

when: sensitive_data_detected
when: confidence < 0.80
when: schema_valid == false or unsupported_claims > 0
when: vendor_risk == high
when: always

It is deliberately not a language: nothing is evaluated here, and a condition that does not read this way is refused, as is one that reads a signal none of the Gate's Guards reports.

Actions

Action	What it decides
proceed	The work goes on.
pause	The work waits for a revision.
retry	The step is run again, possibly by another Cog.
additional_validation	More Guards are run before deciding.
human_review_required	A person reviews the output before it is used.
human_approval_required	A person approves before the action is taken.
expert_review_required	A named expert reviews.
stop	The Op ends.
stop_and_escalate	The Op ends and somebody is told.

The catalogue

Id	Stage	When	Then
configuration-check	preflight	a Frame is missing, a permission denied or a source unauthorized	stop_and_escalate
sensitive-data-stop	in_flight	sensitive_data_detected	stop_and_escalate
tool-violation-retry	in_flight	tool_violation	retry (twice)
low-confidence-review	post_run	confidence < 0.80	human_review_required
unsupported-claims-revision	post_run	schema_valid == false or unsupported_claims > 0	pause
consensus-disagreement-review	post_run	consensus_disagreement > 0.25	expert_review_required
release-approval	post_run	always	human_approval_required
quality-drift-review	continuous	regression_rate > 0.10 or outcome_met == false	expert_review_required

In Python

from agentspecs.gates import get_gate, signals_of

gate = get_gate("low-confidence-review")
gate.then, gate.signals             # GateAction.HUMAN_REVIEW_REQUIRED, ['confidence']
signals_of("schema_valid == false or unsupported_claims > 0")