Agent Infrastructure

ArcTrigger

FastAPI SSE bridge that loads agent identity packs, calls Claude API with full streaming, and relays Server-Sent Events to Surface in real time. The execution host for all named Grid agents. Live on ArcHelm at port 8042 since 2026-08-22.

FastAPISSEport 8042ArcHelmLive

Overview

ArcTrigger is the execution host for Grid agents. When Surface (or any caller) wants to run an agent, it sends a request to ArcTrigger. ArcTrigger loads the agent's identity pack, calls Claude API with the pack's system prompt and tools, and streams the response back over SSE.

It lives at ~/arc-trigger/ on ArcHelm and runs as a persistent service.

Live since 2026-08-22

ArcTrigger is complete and running on ArcHelm. All named Grid agents (Trebek, Donut, Scout, Argus, Ghost, Guardian, Raye, Dax, Louis, Odo, Aria, Locutus) are executed via ArcTrigger identity packs.

Identity packs

An identity pack is a JSON file that defines everything about an agent's execution context:

{ "agent_id": "trebek", "system_prompt": "You are Trebek, the Knowledge & Interface Agent...", "model": "claude-sonnet-4-6", "tools": ["search_knowledge_graph", "recall_memory", "write_memory", "find_entity"], "mcp_servers": ["neo4j-memory", "atlas-memory"], "memory_namespace": "trebek", "charter_scope": "routing" }

Packs are stored in ~/arc-trigger/packs/ on ArcHelm. Changing an agent's behavior means editing its pack — not redeploying ArcTrigger itself.

Request flow

  • Surface sends POST /trigger/{agent_id} with the user message
  • ArcTrigger loads packs/{agent_id}.json
  • Calls Claude API via ArcRouter with the pack's system prompt, tools, and MCP servers
  • Streams each SSE event chunk back to Surface as it arrives
  • Surface renders the streaming response in the agent pane

Agent tools

Tools available to agents running via ArcTrigger (defined in the identity pack's tools array):

search_knowledge_graph(query)
Semantic search over the Grid's neo4j knowledge graph via neo4j-memory MCP.
recall_memory(query, tier?)
Retrieve agent-specific memories from atlas-memory MCP using RRF hybrid search.
write_memory(tier, content, tags[])
Write to the agent's memory namespace in atlas-memory MCP.
find_entity(name)
Look up a specific entity in the neo4j knowledge graph by name.

Endpoints

# Trigger an agent (streaming SSE response) POST http://archelm:8042/trigger/{agent_id} Content-Type: application/json { "message": "...", "session_id": "..." } # List available agents GET http://archelm:8042/agents # Health check GET http://archelm:8042/health

Adding a new agent

  • Create ~/arc-trigger/packs/{agent_id}.json with the agent's system prompt, tools, and MCP config
  • ArcTrigger hot-reloads packs — no restart required
  • Add the agent to Surface's agent registry to make it accessible in the UI
  • Add an identity page to the internal hub if the agent has a defined service scope