Swan GTM skills architecture
Swan's GTM skills architecture splits agent knowledge into foundations and maps. Foundations hold reusable GTM judgment. Maps are thin workflow routes that open the right foundations in the right order.
The architecture at a glance
tl;drThe useful lesson is not the exact Swan playbook. It is the separation of concerns: foundations own reusable judgment, while maps own workflow order. When ICP, scoring, routing, or voice changes, the update happens once in a foundation and every map that reads it benefits.
Before: context embedded in every workflow
One ICP change becomes many edits, and drift is hard to spot.
After: workflows point into shared foundations
One foundation change updates every map that reads it.
Foundations hold the how
exploreA foundation is reusable context that should not belong to any one play. Pick a foundation to see what it owns, its public sub-skills from Swan's playbook, and which maps read it.
CRM
Lookup, dedupe, ownership, deal state, data quality, and write permissions.
How agents read, trust, update, and avoid duplicating CRM records.
Every agent that touches a person, company, deal, or signal should use one shared CRM contract.
Maps hold the where
interactiveA map is one workflow written as a route. It carries little logic of its own. It declares the trigger, then opens the foundations needed for that GTM moment.
Inbound demo
Check whether the account is known, brief the rep, qualify, score, route, draft, and alert.
High-intent inbound should route through every shared judgment layer.
Inbound demo as a route
walkthroughSwan's public playbook shows inbound demo as the clearest map: seven stops, each opening a foundation sub-skill. Click a stop to see what the agent is loading at that moment.
CRM /lookup
Do we know this account?
Where logic should live
decisionThe rule is strict: if the logic repeats across workflows, it belongs in a foundation. If it only defines one workflow's order or trigger, it belongs in a map.
Who is good?
ICP fit, exclusion rules, segment caveats, and buyer roles belong in reusable context.
Borrowing this for AgencyCore
applyWe can adapt the pattern into our own agent context architecture: shared foundations for GTM judgment, thin maps for workflows, and evals around both.
Inventory duplicated context
Find where ICP, scoring, routing, voice, and alert payloads are repeated across AgencyCore workflows.
Good first foundation
Start with CRM because every GTM map needs to know whether a person, company, deal, or signal already exists. That gives agents a shared read/write contract before they act.
Good first map
Use an inbound lead or visitor reveal map first. The trigger is concrete, the route is short, and the human review gate can stay close to the rep.
Sources and caveats
researchThis page is grounded in Swan's public post, public site, and public playbook bundle. The original post mentions Positioning, but the public architecture renders seven foundations without a separate Positioning block, so this page treats Positioning as folded into ICP, Research, or Outreach Voice.