HALO

✨ RLM-based Automatic Agent Optimization Loop ✨

What is this? • Install • Why RLM? • Benchmarks • Development • Contributing

What is this?

Note: If you're looking for a hosted, plug-and-play version of HALO, please sign up for inference.net.

HALO (Hierarchical Agent Loop Optimization) is a methodology for building recursively self-improving agent harnesses using RLMs. This repository contains:

• Information on HALO methodology.
• A Python package that implements the core HALO-RLM engine. View on PyPI
• A demo project that shows how to build HALO loops for your agents using the Python package. View demo
• Benchmarking examples applying HALO to popular agent benchmarks. (View AppWorld).

HALO Loop

The core HALO loop is surprisingly simple:

1. Collect execution traces from your agent harness. HALO uses OpenTelemetry-compatible tracing.
2. Feed traces into HALO-RLM engine.
3. The engine decomposes the traces to understand common failure modes across harness executions and produces a report with its findings.
4. This report is fed into a coding agent like Cursor or Claude Code to generate and apply a set of changes to your harness.
5. The harness is then re-deployed, more traces are gathered, and the cycle repeats.

HALO is great at finding issues in production agent deployments. We find high-traffic environments tend to generate more data with higher variance across executions, creating the type of issues that HALO is great at identifying.

Why an RLM?

A general-purpose harness like Claude Code is the wrong tool for trace analysis. This isn’t because the model isn’t smart, but because traces can get extremely long, and you need a specialized toolkit in order to make observations about systemic agentic behavior. We noticed in our testing that harnesses like CC would often overfit to an error present in a single/few traces rather than generalize to harness-level problems. This led us to creating a specialized form of a RLM.

Get Started

Install

Install the HALO engine + CLI from PyPI:

pip install halo-engine

# Verify installation
halo --help

Usage

1. Integrate Tracing
2. Collect traces by running your agent
3. Run the HALO engine

export OPENAI_API_KEY=...
# Optional: point HALO at another OpenAI-compatible provider.
export OPENAI_BASE_URL=https://openrouter.ai/api/v1

halo path_to_your_traces.jsonl -p "Diagnose errors you find and suggest fixes"

HALO uses the canonical OpenAI env vars: OPENAI_API_KEY for credentials and OPENAI_BASE_URL for OpenAI-compatible providers. If OPENAI_BASE_URL is unset, HALO uses https://api.openai.com/v1. Run halo --help to see all CLI options. The CLI mirrors the model/provider settings exposed by the Python SDK's ModelConfig and ModelProviderConfig.

CLI options

Flag	Default	Description
TRACE_PATH	required	JSONL trace file
--prompt, -p	required	User prompt sent to the root agent
--model, -m	gpt-5.4-mini	Model name for root, subagent, synthesis, and compaction calls
--max-depth	2	Max subagent recursion depth
--max-turns	20	Max turns per agent
--max-parallel	10	Max concurrent subagents
--base-url	OPENAI_BASE_URL / https://api.openai.com/v1	OpenAI-compatible API base URL
--api-key	OPENAI_API_KEY	Provider API key
--header, -H	unset	Provider header as NAME: VALUE. Repeat for multiple headers, matching curl's -H convention
--temperature	provider default	Sampling temperature forwarded to the model
--max-output-tokens	provider default	Maximum output tokens forwarded to the model
--parallel-tool-calls / --no-parallel-tool-calls	enabled	Allow models to issue parallel tool calls
--refusal-retries	0	Retry an agent model request this many times when the model refuses
--reasoning-effort	model/provider default	Reasoning effort for root, subagent, and synthesis calls. Compaction never uses reasoning
--telemetry	off	Emit OpenInference traces of HALO's own LLM, tool, and agent activity

For example:

halo path_to_your_traces.jsonl \
  -p "Diagnose errors you find and suggest fixes" \
  --base-url https://openrouter.ai/api/v1 \
  -H "HTTP-Referer: https://example.com"

Telemetry

HALO can emit OpenInference-shaped traces of its own LLM, tool, and agent activity. It is off by default; nothing is emitted unless you pass --telemetry.

halo TRACE_PATH --prompt "..." --telemetry

When telemetry is enabled, CATALYST_OTLP_TOKEN uploads spans to inference.net Catalyst over OTLP. If it is unset, spans are written to a local JSONL file at ./halo-telemetry-{run_id}.jsonl in the current working directory.

Var	Default	Purpose
CATALYST_OTLP_TOKEN	unset	If set, uploads to Catalyst over OTLP. If unset, writes JSONL locally
CATALYST_OTLP_ENDPOINT	catalyst-tracing default	OTLP endpoint base URL, for example https://telemetry.inference.net
CATALYST_DEBUG	unset	Set to 1 to surface OTLP export errors at WARNING level
CATALYST_TRACING_RUN_ID	unset	Uses this HALO run id instead of a generated uuid
CATALYST_TRACING_*	unset	Generic catalyst-tracing passthrough
HALO_TELEMETRY_PATH	./halo-telemetry-{run_id}.jsonl	Local fallback file path. Only used when CATALYST_OTLP_TOKEN is unset

We have provided a simple demo and an AppWorld demo.

Python API

The engine exposes four entry points from engine.main. Use whichever matches the trade-off you want between observability and code simplicity. The yielded types (AgentOutputItem and AgentTextDelta) are defined in engine/models/engine_output.py:

Function	Sync / async	Returns	When to use
stream_engine_async	async	AsyncIterator[AgentOutputItem | AgentTextDelta]	You want every event including streaming-token deltas (live UI, custom rendering).
stream_engine_output_async	async	AsyncIterator[AgentOutputItem]	You want to log / persist each completed step (assistant message, tool call, tool result) as it lands.
run_engine_async	async	list[AgentOutputItem]	You want the final list at the end and don't care about per-step observability.
stream_engine	sync	Iterator[AgentOutputItem | AgentTextDelta]	Sync generator; yields every event including deltas. Drives the async iterator on a private event loop.
stream_engine_output	sync	Iterator[AgentOutputItem]	Sync generator; yields completed items only. Same shape as the async variant for sync callers.
run_engine	sync	list[AgentOutputItem]	Sync, collects to a list. Pure convenience over asyncio.run(run_engine_async(...)).

from engine.main import stream_engine_output_async

async for item in stream_engine_output_async(messages, cfg, trace_path):
    logger.info("step", extra={"sequence": item.sequence, "agent": item.agent_name})
    # item.item is an AgentMessage (assistant / tool / etc.)

Benchmarks

HALO is consistently capable of driving improvements on benchmarks, solely by optimizing the harness.

AppWorld

We applied HALO to the AppWorld benchmark, a set of agentic tasks that assess the LLM’s ability to use multi-app services like Spotify, Venmo, file systems, and phone contacts. We tested HALO’s ability to improve harnesses for both Gemini 3 Flash and Sonnet 4.6. We iterated on the harness using the dev split, and then used the test_normal split as a proxy to verify that improvements did not come from overfitting.

The feedback from HALO Engine surfaced failures in the harnesses such as hallucinated tool calls, redundant arguments in tools, refusal loops, and semantic correctness issues. Each issue mapped cleanly to a direct prompt edit. HALO’s claims were independently verified from the source trace files with the findings holding up under scrutiny.

The peak improvements over baseline were substantial for both models. For Gemini 3 Flash, dev SGC went from 36.8% to 52.6% (+15.8 points) and test_normal SGC went from 37.5% to 48.2% (+10.7 points). For Sonnet 4.6, dev SGC went from 73.7% to 89.5% (+15.8 points) and test_normal SGC went from 62.5% to 73.2% (+10.7 points).

Development

Local development against this repo uses uv for dependency management and go-task as the task runner.

Setup

git clone https://github.com/context-labs/HALO
cd HALO
task env:setup

task env:setup installs uv (if missing), syncs the venv from uv.lock, and configures the repo's git hooks. After that, the halo CLI is available via uv run halo ... (or activate .venv/).

Common tasks

Run task --list for the full list. The ones you'll use most:

Task	What it does
task check	Run all pre-commit checks: pinned-versions, lint, format, typecheck, unit tests
task check:fix	Same, but auto-fix lint/format issues
task test:unit	Unit tests under tests/unit/
task test:integration	Integration tests under tests/integration/

License

MIT

Contributing

Contributions are welcome! Please feel free to submit a pull request.