II.
AgentCoreImpl overview
Reference · liveagent-core-impl:a5c.core@current
a5c Unified Core (current) — unified agent loop overview
Inspect the raw attributes, linked wiki pages, and inbound or outbound graph edges for agent-core-impl:a5c.core@current.
Attributes
displayName
a5c Unified Core (current) — unified agent loop
agentVersionId
packageRef
loopIteratorPolicy
custom
loopIteratorNotes
a5c's core is modeled as a native compatibility layer for agent loops, not
as a wrapper built on top of Pi or any single backend. It can present the
Pi-compatible serialized tool-use loop while also normalizing Claude Code,
Codex, Gemini, OpenCode, LangGraph/custom graph, and adapters style loop
semantics behind the same AgentCoreImpl contract.
contextManagementStrategy
platform-managed
compactionTriggerNotes
a5c owns the compatibility contract for context and compaction metadata at
the platform boundary. Individual backends may use provider-managed,
rolling-summary, server-managed, or user-managed context; a5c maps those
postures into one platform-managed profile so legacy Pi-style definitions
and newer graph/custom-agent definitions remain comparable.
subagentInvokerPolicy
platform-mediated
subagentInvokerNotes
Subagent dispatch is mediated by the a5c platform contract. Pi-style
tool-call handoff, Claude-style subagents, LangGraph subgraphs/supervisors,
and skill-as-tool execution all normalize to platform-mediated child work
without making a5c depend on any one mechanism.
resultSynthesisPolicy
post-processed
resultSynthesisNotes
a5c synthesizes a normalized result envelope from backend-native terminal
messages, graph terminal state, streamed text, tool results, and generated
artifacts. Pi's model-direct final text is a supported profile, not the
base implementation.
stopDetectionStrategy
structured-end-event
transportClientLibrary
a5c-unified-transport
supportedTransportProtocols
parallelToolCallHandling
native
programmaticToolCalling
optional
programmaticToolCallingNotes
a5c agent-core can optionally expose `code_executor`, a Code Mode /
Programmatic Tool Calling surface inspired by UTCP Code-Mode and Open-PTC.
The executor runs a bounded JavaScript tool chain against the configured
agent-core tools so a model can batch discovery, schema fetches, filesystem,
shell, web, and host tools behind one model-level call.
streamingFidelity
full
thinkingChannelHandling
model-dependent
streamingBlocks
- text
- thinking
- tool_use
effortLevelDelegation
passthrough-to-harness
notes
This node is intentionally the Layer-4 unified contract for a5c. It is
capability-compatible with the legacy Pi wrapper's core definition
(Anthropic/OpenAI transports, tool-use loop, structured end events,
sequential tool handling) while also covering same-layer alternatives that
expose parallel tool calls, graph-defined loops, server-managed loops, or
provider-specific thinking channels. Pi is a compatibility target, not an implementation base. Compatibility is modeled as a native a5c contract; there is no depends_on edge to Pi.
Outgoing edges
composes1
- agent:a5c·AgentProducta5c
depends_on1
- source-ref:a5c-cli·SourceRefa5c CLI
fires_hook11
- hook-surface:phase-change·HookSurfacePhaseChange (a5c canonical)
- hook-surface:phase-change-check·HookSurfacePhaseChangeCheck (a5c canonical)
- hook-surface:wake·HookSurfaceWake (a5c canonical)
- hook-surface:decision-point·HookSurfaceDecisionPoint (a5c canonical)
- hook-surface:start·HookSurfaceStart (a5c canonical)
- hook-surface:done·HookSurfaceDone (a5c canonical)
- hook-surface:session-start·HookSurfaceSessionStart (canonical)
- hook-surface:session-end·HookSurfaceSessionEnd (canonical)
- hook-surface:stop·HookSurfaceStop (canonical)
- hook-surface:pre-tool-use·HookSurfacePreToolUse (canonical)
- hook-surface:post-tool-use·HookSurfacePostToolUse (canonical)
realizes1
- layer:4-agent-core·LayerAgent-Core
speaks7
- model-transport:anthropic-messages·ModelTransportProtocolAnthropic Messages
- model-transport:openai-responses·ModelTransportProtocolOpenAI Responses
- model-transport:openai-chat-completions·ModelTransportProtocolOpenAI Chat Completions
- model-transport:gemini-generate-content·ModelTransportProtocolGemini generateContent
- model-transport:openai-compat·ModelTransportProtocolOpenAI-compatible chat completions
- model-transport:openrouter·ModelTransportProtocolOpenRouter
- model-transport:github-copilot·ModelTransportProtocolGitHub Copilot
supports6
- capability:supports-tool-use·CapabilitySupports tool use
- capability:parallel-tool-calls·CapabilityParallel tool calls
- capability:programmatic-tool-calling·CapabilityProgrammatic tool calling
- capability:streaming·CapabilityStreaming
- capability:can-multi-turn·CapabilityCan multi-turn
- capability:supports-thinking·CapabilitySupports thinking
supports_orchestration_primitive3
- orch-primitive:agent-selection·OrchestrationPrimitiveAgent Selection
- orch-primitive:context-injection·OrchestrationPrimitiveContext Injection
- orch-primitive:telemetry-collection·OrchestrationPrimitiveTelemetry Collection
Incoming edges
about_subject1
composed_of1
- agent-version:a5c@current·AgentVersion