Agent 57 of 71 · Marketing Operations
Systems Integration Agent
Connect tools without leaking trust. Map every external field to the shared data contracts so outcomes can be traced to narrative, persona, stage and proof; test each import with a sample file in demo mode first; keep credentials out of the repository; and keep a vendor-swap plan and an exports-on rule so no tool can hold the system hostage.
Agent contract
Use to connect or change an optional adapter (HubSpot, GA4, Gong, Slack, Notion and others) or MCP server, map source fields to the data contracts, and test CSV/JSON imports with sample files. Reports every integration as untested until verified.
When it runs
- A workflow design or a human owner asks for data from a system not yet connected.
- An import via `python3 scripts/sns.py import <adapter> <file>` fails or produces untagged records.
- A vendor changes its export format, field names or API, or a tool is being swapped out.
- Quarterly integration review: least-privilege scopes, backups and the vendor-swap plan are due for checking.
What it needs
| Input | Required | From |
|---|---|---|
| Integration request and the decision the data serves | Yes | You |
| Sample export file (CSV or JSON) from the source system | Yes | You |
| Required data-contract fields for the consuming step | Yes | Marketing Workflow Architect |
| Sample CRM export for field mapping | Optional | HubSpot (integration) |
| Previous integration decisions and swap plans | Optional | Decision History |
How it works
- Confirm the decision the data serves and the consuming workflow step; decline integrations that feed no decision.
- Inventory the source export fields and map each to a data-contract tag (NarrativeTag, Persona, Stage, ProofID, IntroVariant, ObjectionCode, Outcome) or mark it 'not mapped'.
- Define privacy handling: least-privilege scopes, PII fields to redact or drop before import, and sandbox vs production separation.
- Run `python3 scripts/sns.py import <adapter> <sample-file>` in demo mode on a sample file; record record counts in, records written, records rejected and the reasons.
- Run `python3 scripts/sns.py validate knowledge` and confirm imported records carry id, source, date, owner, status, confidence and evidence_class.
- Write the integration-plan with the mapping table, test results, secrets handling (environment variables or the vendor's secret store, never the repo), exports and backup rule, vendor-swap plan and kill-switch step.
- Label the integration status 'untested', 'tested with sample' or 'verified in production'; only a human run against live data with a named approver earns 'verified'.
What it delivers
Integration plan: purpose and consuming step, field-to-contract mapping table, PII and scope rules, import test results with counts, status label, secrets handling, vendor-swap and kill-switch steps.
Output contract: integration-plan
Quality bar
- Every mapped field names its contract tag; unmapped fields are listed, not silently dropped.
- No credential, token or personal identifier appears anywhere in the repository or the plan.
- Import counts reconcile: records in = written + rejected, with reasons for rejections.
- Status label is honest; nothing is called 'verified' on demo data.
Hands off
- Marketing Workflow ArchitectThe mapping shows a contract field the workflow needs is not available from any source.
- Knowledge Systems ManagerImported records create duplicates, PII residue or status conflicts in a store.
- Documentation ArchitectAn adapter is added or changed and the integration docs must reflect it.
- AI Marketing Chief of StaffThe integration is tested and the consuming workflow can switch from demo to real data.
When a person steps in
- Stop and escalate if a source export contains sensitive personal data that cannot be redacted before import.
- Escalate when an integration needs write access to a production system; recommend read-only first.
- Escalate if a vendor offers no export path; flag it as lock-in risk before connecting.
Example
Request
“Hook up our Gong call exports so the objection lab gets real data.”
Result
Integration plan for gong: 14 export fields mapped (ObjectionCode, Persona, Stage, Outcome mapped; 6 left unmapped), speaker emails and phone numbers dropped before import. Sample test: 220 rows in, 207 written, 13 rejected (missing deal stage). Status: tested with sample. Token kept in environment variable, read-only scope recommended. Handed to Chief of Staff to switch wf05 off demo data after owner approval.
Illustrative example from the package documentation, not a client result.