Skip to content
LocoLLM LocoLLM

LocoAgente as the Multi-Turn Harness

LocoLLM’s Ollama-served models can serve as the inference backend for LocoAgente, the conversational harness library that handles multi-turn orchestration patterns — Debate, Perspective, Synthesis, Iterative Refinement.

The integration in one paragraph

Section titled “The integration in one paragraph”

LocoAgente’s Inference subsystem speaks OpenAI-compatible HTTP. Ollama exposes an OpenAI-compatible endpoint (see ADR-0006). Pointing LocoAgente at that endpoint is a config change, not a code change. LocoLLM’s adapter routing and specialist selection happen on this side of the boundary; LocoAgente sees a model that responds to chat completions and treats it like any other backend.

What composes

Section titled “What composes”
LocoLLM providesLocoAgente provides
Base model + adapter routing (single-turn specialisation)Multi-turn orchestration patterns
Inference engine choice (Ollama; ADR-0006)OpenAI-compatible client (harness/inference/client.py)
Adapter training pipelineContext, Tools, Conversation subsystems
Single-call scaffolding studiesMulti-call scaffolding studies

Why this matters

Section titled “Why this matters”

Frontier-equivalent multi-turn capability on local hardware needs both sides:

  • Specialised models — LocoLLM closes single-call quality gaps via adapters and prompting techniques.
  • Multi-turn harness — LocoAgente gets reliable work out of those models across debate, refinement, and verification loops.

Used alone, each is half the system. Composed, they answer the lab’s central question — how do we get useful work out of small local models? — at both the per-call and per-conversation layers.

Configuration example

Section titled “Configuration example”
# In a LocoAgente harness:
from harness.inference.client import OpenAICompatibleClient


client = OpenAICompatibleClient(
    base_url="http://localhost:11434/v1",
    api_key="ollama",
    model="qwen3-4b-locollm-router",  # the LocoLLM-served model name
)

That is the entire integration. Every LocoAgente Orchestration pattern (SinglePass, DebatePattern, SynthesisPattern, IterativeRefinement) accepts any InferenceClient — so a LocoLLM-served model fits all patterns by construction.

Boundaries

Section titled “Boundaries”

The OpenAI-compatible chat-completions contract is the entire interface. From LocoLLM’s side:

  • LocoLLM does not see Orchestration state or conversation history — it sees individual prompt/completion pairs, one per harness call.
  • LocoLLM does not know which Orchestration pattern is active; the same model serves a SinglePass request and a Debate-round request identically.
  • Adapter routing inside LocoLLM is unaffected by the harness’s structure — each call is routed on its own merits.

This boundary is what lets each side evolve independently. New orchestration patterns in LocoAgente, or new adapters in LocoLLM, do not require changes on the other side.

Further reading

Section titled “Further reading”