Agents

Agents model an LLM reasoning loop as a first-class BLOGE extension. The bloge-agent-ext module adds an agent DSL, Java builder, synchronous and streaming loop operators, tool dispatch, memory strategies, and trace integration.

Core Concepts

Concept	Meaning
Agent	LLM loop definition with model, prompt, memory, tools, and exit condition
Tool	BLOGE operator-backed capability callable by the model
Memory	Retention strategy for messages between turns
Exit condition	Expression deciding when the loop stops
Streaming agent	Agent variant that emits incremental chunks

DSL Shape

agent customerSupport {
  model                = "gpt-6-astra"
  system_prompt        = "You are a helpful customer support agent."
  max_turns            = 6
  max_tool_concurrency = 2
  temperature          = 0.2
  memory               = sliding_window(20)

  tool searchKnowledgeBase : KBSearchOperator {
    description = "Search the knowledge base for a relevant article"
    input {
      query = tool_args.query
    }
  }

  exit_condition = finish_reason == "stop" || tool_call("escalateToHuman")
}

Runtime Contract

AgentLoopOperator expects an llmChat operator in the hosting registry. Each declared tool runs as its own compiled BLOGE graph, so tool execution reuses normal graph semantics rather than a separate tool runtime.

StreamingAgentLoopOperator expects llmStreamingChat and emits AgentStreamChunk values through the streaming channel.

Memory Strategies

DSL form	Behavior
full()	Keep the whole conversation
sliding_window(n)	Retain recent messages
token_budget(n)	Retain as much history as fits a token budget
summary(n)	Summarize older history when it exceeds the limit

Tool Guidance

• Keep tools narrow and observable.
• Give tools explicit descriptions; they become part of the model-facing contract.
• Use max_tool_concurrency when tools can fan out.
• Treat tool input schemas as public to the model.
• Run expensive or isolated tools as Remote Workers when needed.

Related AI Surface

Agent runtime is different from AI authoring. AI authoring generates or validates .bloge source; agents execute LLM loops inside a graph. See AI Agents & LLM Operators for the broader AI operator and authoring overview.