Agentic AI Atlasby a5c.ai
OverviewWikiGraphFor AgentsEdgesSearchWorkspace
/
GitHubDocsDiscord
iiRecord
Agentic AI Atlas · Hermes AIAgent Loop
agent-core-impl:hermes.agent-loopa5c.ai
Search record views/
Record · tabs

Available views

II.Record viewspp. 1 - 1
overviewjsongraph
II.
AgentCoreImpl overview

agent-core-impl:hermes.agent-loop

Reference · live

Hermes AIAgent Loop overview

The AIAgent class in run_agent.py (~4,400 lines) is the single monolithic core loop powering all Hermes surfaces (CLI, gateway, ACP, batch, API). Implements synchronous orchestration with interruptible API calls via background threads, sequential or concurrent tool dispatch, prompt caching, context compression, and subagent delegation.

AgentCoreImplOutgoing · 2Incoming · 0

Attributes

displayName
Hermes AIAgent Loop
agentVersionId
agentVersion:hermes:ge-0-0-0
loopKind
synchronous-iterative
description
The AIAgent class in run_agent.py (~4,400 lines) is the single monolithic core loop powering all Hermes surfaces (CLI, gateway, ACP, batch, API). Implements synchronous orchestration with interruptible API calls via background threads, sequential or concurrent tool dispatch, prompt caching, context compression, and subagent delegation.
keyFiles
  • run_agent.py
  • agent/prompt_builder.py
  • agent/context_engine.py
  • agent/context_compressor.py
  • agent/prompt_caching.py
  • agent/auxiliary_client.py
  • model_tools.py
turnLifecycle
  • Generate task_id
  • Append user message
  • Build or reuse cached system prompt
  • Preflight compression check (>50% context)
  • Build API messages
  • Inject ephemeral prompt layers
  • Apply prompt caching markers (Anthropic)
  • Make interruptible API call
  • Parse response: tool calls loop, text response returns
toolDispatch
Central registry, auto-discovery at import. Sequential for single, ThreadPoolExecutor for multiple.
budgetDefault
90
subagentBudgetDefault
50
compressionThreshold
50% preflight, 85% gateway
compressionStrategy
Memory flush, middle turn summarization, last N preserved, tool call/result pairs kept
apiModes
  • chat_completions
  • codex_responses
  • anthropic_messages
ourEquivalent
Our agent core lives in the upstream harness (Claude Code, Codex, etc.). genty-platform wraps harness cores via adapter bridge (packages/genty/platform/src/harness/adapters/). The babysitter SDK drives turn lifecycle through task definitions and effects rather than owning the inference loop directly.

Outgoing edges

part_of1
  • agent:hermes·AgentProductHermes Agent
realizes1
  • layer:4-agent-core·LayerAgent-Core

Incoming edges

None.

Related pages

No related wiki pages for this record.

Shortcuts

Open in graph
Browse node kind