Implement LLM request batching for concurrent agent decisions

[justinmadison]

Problem

Currently, each agent makes an individual LLM call during AgentRuntime.process_tick(). While agents are processed concurrently via ThreadPoolExecutor, their LLM requests are not batched together, which underutilizes vLLM's continuous batching capabilities.

Current Flow

# python/agent_runtime/runtime.py:87
for agent_id, agent in self.agents.items():
    task = asyncio.create_task(self._agent_decide(agent))
    # Each agent makes individual LLM call

Proposed Solution

Implement batch LLM generation in AgentRuntime.process_tick():

1. Collect all agent contexts first:

   contexts = {}
   for agent_id, agent in self.agents.items():
       contexts[agent_id] = agent._build_context()

2. Send all prompts to vLLM together:

   if isinstance(self.backend, VLLMBackend):
       results = await self.backend.generate_batch(prompts)
   else:
       # Fallback to concurrent individual calls
       results = await self._concurrent_llm_calls(contexts)

3. Parse results into actions:

   actions = {}
   for agent_id, result in results.items():
       actions[agent_id] = agent._parse_action(result)

Expected Impact

• 50-70% faster LLM inference with 4+ agents
• Better utilization of vLLM's continuous batching (PagedAttention)
• With 4 agents: ~1.5x time of single agent instead of 4x

Files to Modify

• python/agent_runtime/runtime.py - Implement batch processing in process_tick()
• python/backends/vllm_backend.py - Add generate_batch() method
• python/backends/base.py - Add abstract generate_batch() interface
• python/backends/llama_cpp_backend.py - Implement fallback (no native batching)

References

• vLLM Continuous Batching: https://docs.vllm.ai/en/latest/serving/engine_args.html
• Current implementation: python/agent_runtime/runtime.py:66-100
• Backend interface: python/backends/base.py:68-86

Priority

HIGH - This is the primary bottleneck in multi-agent scenarios

[justinmadison]

Closing: Superseded by framework integration strategy. LLM calls are now owned by the agent framework (LangGraph, Claude SDK, etc.), not our code. Frameworks handle batching and concurrency internally. See #74 and docs/framework_integration_strategy.md.