CogitaveLearn

The agent fleet

Two named catalogs implement the contract from the previous unit. Both are Day-0 specs, paired as a human-authoritative .md and (for the scheduled catalog) a machine-authoritative .yaml, catalogued in agents/README.md and agents/operations/README.md.

The five scheduled agents (engineering)

agents/scheduled/ runs the engineering estate. Each one is read-mostly or propose-only and never merges, releases, or edits main:

AgentPurposeNever does
pr-triageClassifies a PR (area, size), applies triage labels, routes it to the right CODEOWNERMerge, approve, or push code
dependency-reviewEvaluates dependency changes for known vulnerabilities, license policy, and provenanceAuto-merge a bump
issue-groomingClassifies issues, requests missing repro/version/logs, detects duplicates, surfaces stale onesClose or delete an issue
changelog-docs-syncDrafts the [Unreleased] changelog entry and flags the docs a change requiresEdit main or bump a version
security-triageClassifies and prioritizes advisories and scan alerts, drafts a triage noteRemediate, merge, or disclose

The six operations agents (business)

agents/operations/ runs the business under the same draft-vs-act rule, but its blast radius is often external and reputational rather than internal:

AgentFunctionDrafts / proposesHuman gate (owner)
content-marketingMarketingContent per the editorial calendarEditor + claims-review; never publishes
seo-geoMarketingSEO/GEO optimizations, rank monitoringMarketing owner approves any change
community-supportSupportTicket triage, draft answers, KB-gap detectionSupport agent approves any customer-facing reply
revops-leadGTM / RevOpsLead enrichment, scoring, CRM hygieneRevOps approves merges; never auto-emails a prospect
finops-anomalyFinance / FinOpsCost-anomaly alerts, spend reports, proposed actionsFinance approves any spend action
status-commsOps / incident responseStatus-page and incident comms draftsIncident commander approves any external post

Why the two catalogs read differently

The operations overview names the gap directly: a scheduled agent's mistake is "mostly internal and reversible - a label, a comment, a PR suggestion." An operations agent's mistake can "mislead a customer, leak PII, overspend, or damage the brand" - so its default posture is PROPOSE-only by construction. No agent in either catalog holds a publish, send, email, or apply capability; those verbs are reserved for the human gate, and an attempt to reach them is a capability-boundary violation, denied and recorded as a security event.

The kill-switch and the eval gate

Every ops agent MUST be independently disable-able in one action, without a deploy, by its owner or an incident responder - triggered by a safety-eval regression, a capability-boundary violation, a confirmed hallucination or brand incident, a data-protection breach, or anomalous behavior. And per the three-part contract, no agent's behavior ships at all until it clears the eval harness on accuracy, coverage, safety, and latency.

NOTE

Adding an agent to either catalog is itself governed: author the paired spec, register it in the AIMS inventory, define its grant and loop placement, and clear the eval gate. Each catalog entry is a stub until that happens - the standard says so explicitly.