Optimize agents from real traces.

Name: HALO
Brand: Inference.net

HALO is an RLM-based open-source AI agent built to identify issues in other AI agents by analyzing OpenTelemetry data. Connect HALO to your favorite coding agent for a complete development flywheel.

View GitHub

Trusted by teams building production AI systems

HALO runs everywhere your agents run.

Run development loops locally, build custom workflows using HALO libraries, and use hosted HALO on inference.net for production.

Desktop

Collect or import OTEL traces and run HALO analysis in the desktop app. Built for local development.

from engine.agents.agent_config import AgentConfig
from engine.engine_config import EngineConfig
from engine.main import stream_engine_output_async
from engine.model_config import ModelConfig
from engine.model_provider_config import ModelProviderConfig
from engine.models.engine_output import AgentOutputItem


@dataclass(frozen=True)
class EngineRunResult:
    """Everything the runtime entrypoint needs after the engine returns.

    ``items`` is the raw, typed pydantic list — callers serialise via
    ``model_dump`` when building the result JSON. Keeping the typed
    list (instead of a pre-serialized ``list[dict]``) means downstream
    code (review tooling, future result shape tweaks) gets full type
    info without re-parsing"

PyPi Package

Build HALO into existing workflows using HALO Engine. Free & Open Source. Works with every agent.

View GitHub

InferenceApp8:20 AM

✓Run completed

Bi-weekly HALO analysis

inf-system / HALO Catalyst

Duration

20m 38s

Schedule

Bi-weekly

Report

## Executive summary

The agent completed most runs successfully, but traces show reliability and observability issues.

- 3 traces dominate latency and cost. Tool usage is highly redundant...

Hosted

Use hosted HALO on inference.net with scheduled reports, Slack integration, and production trace storage.

Install SDK

Add traces, run reports, and ship the fix.

Use HALO to build a powerful development loop from production agent trace data. Run reports, investigate traces with timeline & session views, and send to Claude or Codex for final implementation.