Agent-led GTM lifecycle

GTM Brain is designed so an AI agent can do the repetitive work while Fabric Platform enforces the rules underneath it. The agent reads a tenant-scoped catalog, proposes or invokes only the actions in its registration, and receives an explicit exception when evidence, policy, or authority is missing. Operators work the exceptions instead of manually shepherding every read and draft.

This page separates the implemented production contract from the intended exception-only HITL posture. Do not give an agent a provider credential or direct database connection to fill a product gap; add the missing capability as a governed action first.

The end-to-end loop

What agents can do today

1. Draft the ICP and campaign

An AI assistant can help turn a commercial thesis into a seller profile, ICP, campaign brief, audience definition, email sequence, and reviewer policy. Those proposals become durable, versioned workspace records. Evidence and proof assets are independently evaluated before their exact revisions can enter an execution manifest.

The agent should treat a missing active ICP, ambiguous offer, unapproved proof, or stale manifest as a normal stop condition. It should not work around the condition by inventing a policy in its prompt. See Configure your ICP and Artifacts.

2. Source accounts, people, and intent

Use each source for the evidence it actually owns:

• Apollo is the primary company and contact discovery adapter. It searches within the active ICP, returns firmographics and current triggers, and attaches a bounded eligible contact.
• Clearbit Visitors supplies reviewed company-level website intent. An administrator exports CSV from Clearbit, previews the bounded U.S.-company import, and confirms it. Person-bearing columns, self-owned domains, suppressed companies, duplicates, and unqualified rows fail closed.
• PostHog supplies an inbound lead only after the person submits contact details and grants marketing consent. Anonymous analytics and session activity never become GTM records.
• External agents may propose signals, notes, and tasks through the registration-filtered action catalog. A normal read/proposal registration starts with executionActions: [].
• Instantly may read at most the deployment-pinned, pre-reviewed lead list for optional evidence. The configured campaign itself is created and managed in Instantly, then its campaign ID is tested in GTM Brain. GTM Brain does not create Instantly campaigns, run SuperSearch, or treat Instantly employee/location fields as ICP proof.

Clearbit does not make a company outreach-eligible by itself. Visitor intent joins the prioritization funnel, while the active ICP, firmographics, suppression state, verified contact, public trigger, and approved campaign still decide whether a draft can be produced.

See Integrations and Website leads and PostHog.

3. Run the governed campaign pass

The Temporal campaign pass runs the sense → remember → judge → draft loop against a pinned campaign version. Stable command and workflow identities make retries resumable. A model recommendation is never the mutation boundary: Fabric Platform validates the action schema, actor, tenant, resource scope, policy result, idempotency binding, and audit event before state changes.

A bounded external agent can request only the campaign and action named in its short-lived grant. Catalog discovery does not grant lasting authority; execution rechecks the registration and tenant policy at the final boundary.

4. Send and follow up

For first touch, the final action gtm.send_approved_outreach_v2 re-derives the recipient, approved draft, campaign manifest, content policy, proof, suppression state, frequency cap, inbox ramp, and launch authorization immediately before reservation and delivery. Production also requires GTM_ENABLE_V2_LIVE_DELIVERY=1, GTM_LIVE_DELIVERY_MODE=governed, an exact current authorization, and natural-person approval. The Campaigns UI creates that authorization through the canonical governed action; it never grants delivery authority on its own.

Replies enter through authenticated Instantly or provider-neutral webhooks. Communications shows the full tenant-scoped thread, sender and recipient evidence, customer context, manual and AI reply composition, review state, archive/restore, and dead-lead handling. A rejected AI reply can be regenerated as a bounded next revision. Archive and dead-lead actions preserve immutable history; they do not silently add suppression.

AI reply delivery has its own deployment gate, GTM_ENABLE_AI_REPLY_DELIVERY. When it is off, the review record remains durable and no reply leaves GTM Brain. When it is on, the approved reply is still rechecked for source identity and suppression before a one-attempt provider call. Unknown delivery outcomes require reconciliation rather than an unsafe retry.

See Reply review and Email and calendar synchronization.

5. Record meetings and learn

The current system records a meeting outcome and can synchronize immutable calendar snapshots through gtm.sync_calendar_event_v2. It does not create Google Calendar, Microsoft 365, Calendly, or Cal.com appointments. Until a governed scheduling adapter is implemented and certified, an agent may prepare the handoff and record confirmed evidence, but a person or the connected provider completes the actual booking.

Reply, meeting, bounce, unsubscribe, and delivery outcomes feed the durable account history and reporting projections. Experiments may measure actual delivery and mature outcomes, but cannot become another send path.

Invisible governance, visible exceptions

“Invisible governance” means operators do not manually re-enter policy at every step. It does not mean the platform skips authorization or hides why an action stopped.

The normal path runs these controls automatically:

• tenant and actor authentication;
• registration, campaign, and action scope;
• schema and immutable-version checks;
• suppression, consent, frequency, and inbox-ramp limits;
• approved offer, proof, content-policy, and recipient binding;
• stable idempotency and durable recovery;
• audit, retention, and provider reconciliation.

Escalate to a human when the platform reports an exception such as missing/stale evidence, conflicting policy, low-confidence judgment, requested authority outside the grant, destructive lifecycle action, unknown delivery outcome, or live_delivery_requires_human. Agents should present the evidence and the narrow decision needed; they should not retry a permanent denial as if it were a network error.

Current posture and target posture

Connect an external agent safely

1. Enroll the runtime through the one-time invite flow.
2. Start with read tools and proposal actions; keep executionActions: [].
3. Inspect the catalog returned for that exact registration.
4. Add only the action, campaign, or skill scope needed for the next bounded job, with an expiry.
5. Treat authorization and policy denials as terminal until an operator changes the underlying state.
6. Keep provider, Lakebase, Temporal, Databricks, and Platform Host credentials out of the agent.

See External agent gateway, Fabric Harness agents, and Governance.