{
  "id": "page:docs-v6-spec-and-roadmap-v6-1-agent-layer-capabilities",
  "_kind": "Page",
  "_file": "wiki/docs/v6-spec-and-roadmap/v6-1/agent-layer-capabilities.md",
  "_cluster": "wiki",
  "attributes": {
    "nodeKind": "Page",
    "sourcePath": "docs/v6-spec-and-roadmap/v6.1/agent-layer-capabilities.md",
    "sourceKind": "repo-docs",
    "title": "Agent Layer Capabilities — What agent-core, agent-runtime, and agent-platform Should Actually Do",
    "displayName": "Agent Layer Capabilities — What agent-core, agent-runtime, and agent-platform Should Actually Do",
    "slug": "docs/v6-spec-and-roadmap/v6-1/agent-layer-capabilities",
    "articlePath": "wiki/docs/v6-spec-and-roadmap/v6.1/agent-layer-capabilities.md",
    "article": "\n# Agent Layer Capabilities — What agent-core, agent-runtime, and agent-platform Should Actually Do\n\nThe renames and restructuring are plumbing. This document defines the **capabilities** each layer package needs to be a super-capable implementation, grounded in the atlas graph node kind definitions.\n\n## Agent-Core (L4) — The Brain\n\n**Graph node kind:** `AgentCoreImpl`\n**Graph attributes:** loopIteratorPolicy, contextManagementStrategy, subagentInvokerPolicy, resultSynthesisPolicy, stopDetectionStrategy, parallelToolCallHandling, streamingFidelity, thinkingChannelHandling\n\n**Current reality:** babysitter-sdk's `runtime/orchestrateIteration.ts` — replay engine + process function execution. Purely orchestration, no intelligence.\n\n**What it should be:**\n\n### Core Loop Engine\n- **Multi-strategy loop iteration:** Not just \"run process function until effect\" — support sequential, concurrent, group-chat, and handoff orchestration patterns (per Microsoft Azure AI agent design patterns)\n- **Context window management:** Automatic compaction, sliding window, priority-based context selection. Currently \"user-managed\" — should offer managed strategies\n- **Thinking channel:** First-class thinking/reasoning token handling — route thinking to observer, compress thinking in context, toggle thinking effort per iteration\n- **Streaming fidelity:** Full streaming pipeline from model through transport through agent core to presentation — not just \"capture stdout\"\n\n### Subagent Orchestration\n- **Agent-as-tool:** Invoke another agent as a tool call, with typed input/output contract\n- **Handoff:** Transfer control from one agent to another with context passing\n- **Group chat:** Multiple agents collaborating on a shared context with turn management\n- **Delegation with oversight:** Parent agent delegates to child, reviews result, can reject and retry\n\n### Replay & Determinism\n- **Stable invocation keys:** (PR #178 — in progress) Deterministic replay across code changes\n- **Checkpoint/restore:** Save and restore execution state for long-running processes\n- **Speculative execution:** Try multiple paths, keep the best result\n- **Time-travel debugging:** Replay to any point in the journal, inspect state\n\n### Context Engineering\n- **Structured context injection:** Process-defined context that gets injected into every prompt\n- **Memory-augmented prompting:** Pull relevant memories from knowledge fabric into context\n- **Tool result summarization:** Compress large tool outputs before injecting into context\n- **Dynamic system prompt:** System prompt evolves based on process phase and accumulated context\n\n---\n\n## Agent-Runtime (L5) — The Host\n\n**Graph node kind:** `AgentRuntimeImpl`\n**Graph attributes:** daemon, observer, MCP-server surfaces\n\n**Current reality:** `agent-platform` — CLI binary with daemon, observer, MCP server, harness bridge\n\n**What it should be:**\n\n### Process Lifecycle\n- **Daemon mode:** Long-running service that manages multiple concurrent runs\n- **Hot reload:** Process definitions can be updated without restarting — next iteration uses new code\n- **Graceful shutdown:** In-flight effects complete, state is checkpointed, can resume later\n- **Health monitoring:** Self-diagnosis, stuck-run detection, automatic recovery\n\n### Session Management\n- **Multi-session:** Multiple agent sessions running concurrently with isolated state\n- **Session persistence:** Sessions survive process restart (currently: filesystem, target: pluggable backends)\n- **Session sharing:** Multiple operators can observe/interact with the same session\n- **Session handoff:** Transfer a session from one runtime to another (local → cloud, dev → CI)\n\n### Resource Management\n- **Token budgets:** Per-run and per-session token limits with early warning\n- **Cost tracking:** Real-time cost accumulation across all model calls in a run\n- **Concurrency limits:** Max parallel effects, max parallel runs, queue overflow policy\n- **Timeout cascade:** Per-effect, per-iteration, per-run timeouts with escalation\n\n### Observability\n- **Structured telemetry:** OpenTelemetry spans for every effect, iteration, model call\n- **Live streaming:** Real-time event stream to observer dashboard, mobile, watch\n- **Audit log:** Immutable record of every decision, approval, and state change\n- **Performance profiling:** Per-effect latency, model response time, replay overhead\n\n### Execution Modes\n- **Local:** Current behavior — runs in the host process\n- **Docker:** Spawn agent in a container with mounted workspace\n- **SSH:** Run agent on a remote machine\n- **Kubernetes:** Submit agent run as a K8s Job with resource requests\n- **Cloud (managed):** Submit to a5c cloud service\n\n---\n\n## Agent-Platform (L6) — The Ecosystem\n\n**Graph node kind:** `AgentPlatformImpl`\n**Graph attributes:** defineTask + plugin/skill registry\n\n**Current reality:** `extensions-adapter` (plugin compiler) + `agent-catalog` (metadata) + agent-platform platform module\n\n**What it should be:**\n\n### Plugin System\n- **Hot-loadable plugins:** Plugins can be added/removed/updated without restart\n- **Plugin sandboxing:** Each plugin runs with declared permissions, can't access unauthorized resources\n- **Plugin marketplace:** Browse, search, install plugins from multiple registries\n- **Plugin versioning:** Semantic versioning with compatibility checks and migration scripts\n- **Plugin composition:** Plugins can depend on and extend other plugins\n\n### Skill System\n- **Skill discovery:** Automatic discovery from local files, installed plugins, remote registries\n- **Skill routing:** Intelligent skill selection based on task type, context, and capabilities\n- **Skill chaining:** Skills can invoke other skills with typed interfaces\n- **Skill marketplace:** Share and distribute skills across teams and organizations\n- **Skill testing:** Built-in test harness for skill development\n\n### Process Library\n- **Curated library:** Pre-built processes for common workflows (TDD, code review, debugging, etc.)\n- **Process composition:** Combine processes with phases, branching, and error handling\n- **Process versioning:** Track process evolution with quality scores per version\n- **Process recommendation:** Suggest appropriate process based on task description and repo context\n\n### Agent Teams\n- **Team composition:** Define teams of agents with roles and responsibilities\n- **Team coordination:** Shared context, handoff protocols, conflict resolution\n- **Team scaling:** Dynamic team size based on workload\n- **Team governance:** Approval chains, escalation paths, authority delegation\n\n### Distribution\n- **Cross-harness compilation:** Single plugin → Claude Code + Codex + Gemini + Copilot + Cursor + Pi + OMP + OpenCode + OpenClaw + Hermes (current: extensions-adapter does this)\n- **Package publishing:** Automated npm publish with staging/production tags\n- **Install verification:** Post-install health check per harness\n- **Telemetry:** Plugin usage analytics, error rates, user satisfaction\n\n---\n\n## Implementation Priority\n\n### Tier 1: Core capabilities that unlock everything else\n\n| Capability | Layer | Why first |\n|-----------|-------|-----------|\n| Multi-strategy loop | L4 | Every other capability depends on a flexible loop |\n| Subagent orchestration (agent-as-tool, handoff) | L4 | Enables team agents and complex workflows |\n| Token budgets + cost tracking | L5 | Without this, agents burn money unchecked |\n| Structured telemetry | L5 | Can't improve what you can't measure |\n| Skill routing | L6 | Makes the 2000+ library skills actually discoverable |\n\n### Tier 2: Capabilities that improve quality\n\n| Capability | Layer | Why |\n|-----------|-------|-----|\n| Context window management | L4 | Long conversations degrade without it |\n| Session handoff | L5 | Local dev → CI → cloud workflow |\n| Plugin sandboxing | L6 | Security prerequisite for marketplace |\n| Process recommendation | L6 | Reduces onboarding friction |\n| Docker/SSH execution | L5 | Sandboxed execution for untrusted code |\n\n### Tier 3: Capabilities that enable scale\n\n| Capability | Layer | Why |\n|-----------|-------|-----|\n| Group chat / team agents | L4+L6 | Multi-agent collaboration |\n| Speculative execution | L4 | Parallel exploration |\n| K8s execution | L5 | Cloud-native scaling |\n| Plugin marketplace | L6 | Ecosystem growth |\n| Memory-augmented prompting | L4+L12 | Long-term learning |\n\n---\n\n## Mapping to Process Phases\n\nThe v6.1-graph-alignment.mjs process should be extended with:\n\n| New Phase | Capabilities | Depends On |\n|-----------|-------------|------------|\n| Phase 6: Core loop upgrade | Multi-strategy loop, subagent orchestration, context management | Phase 1.3 (adapters decomposition) |\n| Phase 7: Runtime hardening | Token budgets, cost tracking, health monitoring, graceful shutdown | Phase 3.1 (lifecycle), Phase 2.1 (tools-adapter) |\n| Phase 8: Platform maturation | Skill routing, process recommendation, plugin sandboxing | Phase 1.1 (extensions-adapter), Phase 1.2 (tasks-adapter) |\n| Phase 9: Execution modes | Docker, SSH, K8s execution | Phase 3.1 (9-state lifecycle) |\n| Phase 10: Team agents | Group chat, delegation, coordination | Phase 6 (core loop) |\n",
    "documents": []
  },
  "outgoingEdges": [],
  "incomingEdges": [
    {
      "from": "page:docs-v6-spec-and-roadmap-v6-1",
      "to": "page:docs-v6-spec-and-roadmap-v6-1-agent-layer-capabilities",
      "kind": "contains_page"
    }
  ]
}