Agentic AI Atlasby a5c.ai
OverviewWikiGraphFor AgentsEdgesSearchWorkspace
/
GitHubDocsDiscord
i.4Wiki
Agentic AI Atlas · Agent Stack Decomposition — agent-platform, agent-core, SDK, and adapters
docs/v6-spec-and-roadmap/v6-1/agent-stack-decompositiona5c.ai
Search the atlas/
Wiki · linked records

Article and nearby pages

I.Current articlepp. 1 - 1
Agent Layer Capabilities — What agent-core, agent-runtime, and agent-platform Should Actually DoGap Inventory — Graph vs ImplementationGraph Alignment Tasks — Make the Repo Match the GraphAdapter Architecture — Graph vs Packages Deep DiveNaming Alignment — Graph Concepts vs Package NamesStack Layer Map — Atlas Graph to Package Mapping
I.
Wiki article

docs/v6-spec-and-roadmap/v6-1/agent-stack-decomposition

Reading · 8 min

Agent Stack Decomposition — agent-platform, agent-core, SDK, and adapters reference

The atlas graph models two AgentProducts (agent-platform, adapters) each decomposed into Core/Runtime/Platform/UI implementations. This document maps the graph's decomposition to the actual packages and identifies what needs to change.

Page nodewiki/docs/v6-spec-and-roadmap/v6.1/agent-stack-decomposition.mdNearby pages · 7Documents · 0

Continue reading

Nearby pages in the same section.

Agent Layer Capabilities — What agent-core, agent-runtime, and agent-platform Should Actually DoGap Inventory — Graph vs ImplementationGraph Alignment Tasks — Make the Repo Match the GraphAdapter Architecture — Graph vs Packages Deep DiveNaming Alignment — Graph Concepts vs Package NamesStack Layer Map — Atlas Graph to Package Mappingv6.1 Priorities — Actionable Work Items

Agent Stack Decomposition — agent-platform, agent-core, SDK, and adapters

The atlas graph models two AgentProducts (agent-platform, adapters) each decomposed into Core/Runtime/Platform/UI implementations. This document maps the graph's decomposition to the actual packages and identifies what needs to change.

Graph Model: Two Agent Products

agent-platform (full-cli-agent, stackScope: full)

The graph decomposes agent:agent-platform into 4 implementation layers:

Graph NodeKindLayerDescription
agent-core-impl:agent-platform.core@currentAgentCoreImplL4SDK-backed CLI loop — delegates to babysitter-sdk process functions
agent-runtime-impl:agent-platform.runtime@currentAgentRuntimeImplL5SDK runtime + daemon — daemon, observer, MCP server surfaces
agent-platform-impl:agent-platform.platform@currentAgentPlatformImplL6defineTask + plugin/skill registry
agent-ui-impl:agent-platform.ui@currentAgentUIImplL11CLI binary presentation

adapters (full-cli-agent + remote)

The graph has agent:adapters and agent:adapters-remote:

Graph NodeKindDescription
agent-core-impl:adapters-remote.core@currentAgentCoreImplTransport-only delegator
presentation:adapters-tuiPresentationTUI
presentation:adapters-cliPresentationCLI
presentation:adapters-webuiPresentationWeb UI
presentation:adapters-mobile-iosPresentationiOS
presentation:adapters-mobile-androidPresentationAndroid
+ 4 more (TV, watch)Presentation...

---

Current Package Reality

Three "core" packages doing different things:

Packagenpm nameWhat it actually does
packages/babysitter-sdk/@a5c-ai/babysitter-sdkOrchestration engine: runs, replay, storage, tasks, hooks, plugins, profiles, session, compression, MCP, CLI commands
packages/genty/core/@a5c-ai/genty-coreThin: agentic tools, background process registry, session binding, deferred tool registry
packages/genty/platform/@a5c-ai/genty-platformFat CLI: wraps SDK + agent-core, adds daemon, observer, governance, harness bridge, cost, interaction

The confusion:

1. **genty-core is tiny** — 6 source files. It's NOT the "core" of anything in the graph sense. The graph's AgentCoreImpl (SDK-backed CLI loop) lives in babysitter-sdk, not genty-core.

2. **babysitter-sdk is the real core** — runtime, replay, storage, tasks, hooks, process context, effect model. This is what the graph calls agent-core-impl:agent-platform.core.

3. **agent-platform is core + runtime + platform + UI** — it exports 14 sub-modules (api, cli, cost, daemon, governance, harness, interaction, observability, runtime, seams, session, storage, tasks). The graph correctly decomposes it into 4 layers but the code is one monolithic package.

---

Target: Graph-Aligned Package Structure

agent-platform decomposition

The graph says agent-platform has Core, Runtime, Platform, and UI. The code should reflect this:

Graph LayerTarget PackageCurrent LocationWhat Moves
L4 AgentCoreImpl@a5c-ai/babysitter-sdk (stays)packages/babysitter-sdk/Nothing — SDK IS the core. agent-core package contents fold into SDK or agent-platform.
L5 AgentRuntimeImpl@a5c-ai/genty-platform (stays, slimmed)packages/genty/platform/Keep: daemon, session, harness bridge, runtime. Remove: things that belong in other layers.
L6 AgentPlatformImpl@a5c-ai/extensions-adapter (renamed from extensions-adapter) + @a5c-ai/agent-catalogpackages/extensions-adapter/, packages/agent-catalog/Plugin compilation, skill discovery, marketplace
L11 AgentUIImpl@a5c-ai/genty-platform CLI entrypackages/genty/platform/src/cli/CLI stays in agent-platform — it's the binary

agent-core package fate

@a5c-ai/genty-core is confusingly named — it's not the agent core, it's a small utility package. Options:

**Option A: Fold into agent-platform** (recommended)

  • Move agentic tools, background process registry, session binding into agent-platform
  • Delete agent-core package
  • Update imports

**Option B: Rename to agent-tools**

  • If it grows to own tool dispatch, rename to @a5c-ai/agent-tools
  • But tools-adapter is the graph concept for this — so it would become part of tools-adapter

**Option C: Fold into tools-adapter**

  • When tools-adapter is created (Phase 2.1), merge agent-core into it
  • agent-core's agentic tools + deferred tool registry are tools-adapter concerns

**Recommendation: Option C** — agent-core becomes the seed of tools-adapter.

tasks-adapter (renamed from tasks-adapter)

Per user direction: tasks-adapter becomes part of tasks-adapter.

The graph's adapter:tasks-adapter description: "The lone live Trust Chain primitive — ProvenBreakpointAnswer signs decision answers with the named Authority of the responder. Bridges every backend."

Current PackageTargetWhat Changes
@a5c-ai/tasks-adapter@a5c-ai/tasks-adapterRename package, keep all functionality (approval routing, cryptographic signing, pluggable backends, MCP)

The tasks-adapter scope is broader than breakpoints — it's the trust chain for ALL task resolutions, not just breakpoints. Future: Linear/GitHub/Slack task backends, task routing policies, multi-responder coordination.

---

Detailed Task List

T1: Dissolve agent-core into tools-adapter seed

TaskEffort
Create packages/tools-adapter/ with seed from agent-coreMedium
Move agenticTools/ from agent-core to tools-adapterSmall
Move deferredToolRegistry.ts from agent-core to tools-adapterSmall
Move backgroundProcessRegistry.ts to agent-platform (runtime concern)Small
Move session.ts to agent-platform (runtime concern)Small
Update all imports from @a5c-ai/genty-core across monorepoMedium
Deprecate @a5c-ai/genty-core on npmSmall
Update graph: remove agent-core SourceRef, add tools-adapter SourceRefSmall

T2: Slim agent-platform to match graph Runtime+UI scope

agent-platform currently exports 14 modules. Per graph, it should own Runtime (L5) and UI (L11):

ModuleCurrentTarget OwnerReason
apiagent-platformagent-platformRuntime API surface
cliagent-platformagent-platformUI (L11)
costagent-platformagent-platformRuntime telemetry
daemonagent-platformagent-platformRuntime (L5)
governanceagent-platformagent-platform or tasks-adapterGovernance could move to tasks-adapter if it's approval-centric
harnessagent-platformagent-platformRuntime harness bridge
interactionagent-platformagent-platformUI (L11)
observabilityagent-platformagent-platformRuntime telemetry
runtimeagent-platformagent-platformCore runtime (L5)
seamsagent-platformagent-platformArchitecture seam contracts
sessionagent-platformagent-platformRuntime session (L5)
storageagent-platformagent-platformRuntime storage
tasksagent-platformagent-platformTask execution

**Result:** agent-platform keeps most modules — they ARE runtime/UI concerns. The governance module is the only candidate for extraction (to tasks-adapter), but only if it's purely about approval policies.

T3: babysitter-sdk layer annotation

The SDK is intentionally monolithic. Don't split it — annotate what graph concepts it implements:

SDK ModuleGraph LayerGraph Node Kind
runtime/L4 (Agent-Core) + L13 (Orchestration)AgentCoreImpl, OrchestrationPrimitive, Run, Phase
storage/L7 (Workspace)Workspace (partial)
tasks/L8 (Execution)Effect, Execution, Invocation
hooks/Cross-cuttingHookSurface (via hooks-adapter)
plugins/L6 (Agent-Platform)Plugin, PluginMarketplace
session/L5 (Agent-Runtime)Session, session-storage-adapter
mcp/Tool-adapter concernToolDescriptor, ToolServer
harness/L5 (Agent-Runtime)AgentRuntimeImpl (adapter detection)
profiles/L12 (Knowledge-Fabric)(user/project knowledge)
compression/L4 (Agent-Core)CompactionPolicy
prompts/L4 (Agent-Core)PromptTemplate
breakpoints/L9 (Sandbox) + L14 (Governance)HumanCheckpoint

**Tasks:**

  • [ ] Add "atlas" field to sdk package.json listing all layers and node kinds
  • [ ] Add module-level JSDoc comments referencing graph layer and node kind
  • [ ] Create docs/sdk-layer-map.md documenting which source directory maps to which layer

T4: adapters internal decomposition

As described in graph-alignment-tasks.md Phase 1.3, split into graph-aligned sub-packages:

CurrentTargetGraph Adapter
adapters/coreadapters/comm (or agent-comm-adapter)adapter:agent-comm-adapter
adapters/cli launch.tsadapters/launchadapter:agent-launch-adapter
adapters/cli install.ts + adapters/adaptersadapters/configadapter:agent-config-adapter
adapters/cli (composition)adapters/cli (thin, wires muxes)No adapter — composition
adapters/gatewayadapters/gatewayNo adapter — presentation
adapters/tuiadapters/tuiNo adapter — presentation
adapters/uiadapters/uiNo adapter — presentation
adapters/webuiadapters/webuiNo adapter — presentation
adapters/observabilityadapters/observabilityNo adapter — cross-cutting

---

Execution Order (combined with graph-alignment-tasks.md)

Code
1. Rename tasks-adapter → tasks-adapter
2. Rename extensions-adapter → extensions-adapter  
3. Dissolve agent-core → tools-adapter seed + agent-platform
   ↓
4. Extract agent-launch-adapter from adapters-cli
5. Extract agent-config-adapter from adapters-cli + adapters-adapters
6. Rename agent-comm-adapter → agent-comm-adapter
   ↓
7. Annotate babysitter-sdk with layer metadata
8. Annotate agent-platform with layer metadata
   ↓
9. Implement tools-adapter: schema translation, dispatch policies
10. Implement agent-launch-adapter: 9-state lifecycle, retry
11. Implement agent-comm-adapter: formalized event schema
12. Complete transport-adapter: codecs
   ↓
13. Update graph: SourceRef nodes, decomposition edges, cluster moves

Final Package Landscape (Target)

PackageGraph AlignmentLayer(s)
@a5c-ai/babysitter-sdkOrchestration engine (annotated, not renamed)L4, L7, L8, L13
@a5c-ai/genty-platformAgentRuntimeImpl + AgentUIImplL5, L11
@a5c-ai/babysitterMetapackage (unchanged)—
@a5c-ai/launch-adapteradapter:agent-launch-adapterL8
@a5c-ai/comm-adapteradapter:agent-comm-adapterL4-L5
@a5c-ai/config-adapteradapter:agent-config-adapterL5-L6
@a5c-ai/adapters-cliComposition CLI (adapters)L10
@a5c-ai/adapters-gatewayRemote APIL6
@a5c-ai/genty-tuiPresentationL11
@a5c-ai/genty-uiShared UIL11
@a5c-ai/genty-web-appWeb UIL11
@a5c-ai/adapters-observabilityTelemetryCross-cutting
@a5c-ai/hooks-adapter-*adapter:hooks-adapterCross-cutting
@a5c-ai/transport-adapteradapter:transport-adapterL3
@a5c-ai/extensions-adapteradapter:extensions-adapter (renamed from extensions-adapter)L6
@a5c-ai/tasks-adapteradapter:tasks-adapter (renamed from tasks-adapter)L9, L14
@a5c-ai/tools-adapteradapter:tools-adapter (new, seeded from agent-core)L8
@a5c-ai/agent-catalogKnowledge catalogL12
@a5c-ai/atlasGraph data + indexerCross-cutting
@a5c-ai/triggers-adapterAutomation triggersCross-cutting
@a5c-ai/cloudDeploymentCross-cutting
@a5c-ai/observer-dashboardObservability UIL11
@a5c-ai/kradle-*Project managementL6

**Dissolved:**

  • @a5c-ai/genty-core → absorbed into @a5c-ai/tools-adapter (tools) + @a5c-ai/genty-platform (session/registry)
  • @a5c-ai/adapters-codecs → absorbed into @a5c-ai/config-adapter
  • @a5c-ai/extensions-adapter → renamed to @a5c-ai/extensions-adapter
  • @a5c-ai/tasks-adapter → renamed to @a5c-ai/tasks-adapter

Trail

Wiki
Babysitter Docs
a5c.ai V6 Spec And Roadmap
v6.1 Specification — Layer-to-Package Gap Analysis

Agent Stack Decomposition — agent-platform, agent-core, SDK, and adapters

Continue reading

Agent Layer Capabilities — What agent-core, agent-runtime, and agent-platform Should Actually Do
Gap Inventory — Graph vs Implementation
Graph Alignment Tasks — Make the Repo Match the Graph
Adapter Architecture — Graph vs Packages Deep Dive
Naming Alignment — Graph Concepts vs Package Names
Stack Layer Map — Atlas Graph to Package Mapping
v6.1 Priorities — Actionable Work Items

Page record

Open node ledger

wiki/docs/v6-spec-and-roadmap/v6.1/agent-stack-decomposition.md

Documents

No documented graph nodes on this page.