---
license: other
tags:
  - code-agents
  - software-engineering
  - llm-evaluation
  - reproducibility
---

# Agent Harness Research

[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.21781711.svg)](https://doi.org/10.5281/zenodo.21781711)

**Author:** Mandeep Sidhu, Independent Researcher

**Release:** 1.0.0 preprint candidate

**Paper:** [Dissecting Repository-Scale Code-Agent Harnesses](output/pdf/dissecting_repository_scale_code_agent_harnesses.pdf)

**Archival DOI:** [10.5281/zenodo.21781711](https://doi.org/10.5281/zenodo.21781711)

This repository is an experimental platform for a scientific study of how
repository-navigation harnesses affect an LLM coding agent on codebases that do
not fit in the model context window.

Study 1 fixes **Qwen3.6-35B-A3B** (35B total parameters, 3B active parameters).
The completed Studies 2--5 evaluate that model, **GPT-OSS-20B**, and
**Qwen3-Coder-30B-A3B-Instruct**,
served locally by LM Studio on port `1234`. Experiments abort if the exposed
model, variant, quantization, context, or reasoning default does not match. The
canonical model is documented by the [Qwen release](https://qwen.ai/blog?id=qwen3.6-35b-a3b)
and [official model repository](https://huggingface.co/Qwen/Qwen3.6-35B-A3B).

LM Studio exposes model discovery through `/api/v1/models` and OpenAI-compatible
inference and tool calling through `/v1/chat/completions`. See the
[LM Studio server documentation](https://lmstudio.ai/docs/developer/core/server),
[model listing documentation](https://lmstudio.ai/docs/developer/rest/list), and
[tool-use documentation](https://lmstudio.ai/docs/developer/openai-compat/tools).

## Completed research program

- 21 immutable harness specifications (`H000` through `H020`)
- Fifteen completed experiment families (`E01`--`E05` and `E07`--`E16`)
- 5,453 audited cells: 965 Study 1, 912 Study 2, 540 Study 3, 210 Study 4,
  and 2,826 Study 5
- Separate robustness scenarios and vector-index backend profiles
- Strict local-model discovery with no silent model fallback
- Append-only JSONL telemetry and deterministic run identities
- Confirmatory statistical analysis with paired exact tests, task bootstrap
  intervals, Holm correction, and a mixed-effects factorial model
- A reproducible research paper with generated figures/tables, checksums, and
  page-by-page rendered-PDF quality assurance

## Powered Study 2

- 60 validated real-change tasks across GitLab Runner, GitLab CLI, and python-gitlab
- Go and Python repositories whose full source exceeds the 65,536-token treatment cap
- Five component harnesses plus controlled Agentless-style and SWE-agent-style baselines
- Qwen3.6-35B-A3B and GPT-OSS-20B under identical local-runtime controls
- 840 deterministic main cells plus 72 preregistered stochastic reliability cells
- Exact paired testing, task-cluster intervals, multiplicity control, and hierarchical analysis
- Primary result: H007 hybrid and H000 lexical both resolved 2/60 Qwen tasks
  (risk difference 0.000, 95% task-bootstrap CI [-0.050, 0.050], exact p=1)
- No secondary contrast survived correction; GPT-OSS produced no accepted
  repair and exposed a strong model/tool-protocol compatibility limitation

## Protocol-normalized Studies 3--4

- E09: 60 tasks x 3 models x 3 edit interfaces = 540 real local-LLM cells
- GPT-OSS exact replacement accepted edits on 27/60 tasks versus 0/60 for raw
  diff (RD +0.450, 95% CI [0.333, 0.583], exact p=1.49e-8)
- Resolution-blind gate: P002 for Qwen3.6; P003 for GPT-OSS and Qwen3-Coder
- E10: 20 fresh tasks x 3 gated models x exact/hybrid/oracle retrieval = 180
- Hybrid resolved 1/60 model-task pairs versus exact 0/60 (RD +0.0167,
  task-cluster CI [0, 0.05], exact p=1); oracle file names resolved 4/60
- E11/E12: 30 prespecified stochastic and context-scarcity sensitivity cells

## Harness-design Study 5

- E13: 60 tasks x 8 lexical/syntax/dense combinations x 3 models = 1,440 cells
- E14: 20 tasks x 3 retrievers x 3 edit interfaces x 3 models = 540 cells
- E15: 20 tasks x 9 graph/query/interface/packing harnesses x 3 models = 540 cells
- E16: six deterministically selected archetypes x 17 fresh tasks x 3 models = 306 cells
- No E13 component main effect survives correction; E14 finds no adjusted
  retrieval-by-action interaction
- Iteration and whole-file packing reliably increase cost without a demonstrated
  applicable-patch benefit
- E16 quality-rank transfer is weak (accepted Spearman rho 0.04; applicable
  rho -0.21), while token-cost rank is stable (rho 0.94)
- Only one E16 cell resolves, so the study reports stage-aware trends and
  Pareto candidates rather than a universal winning harness

## Repository layout

```text
configs/
  harnesses/     one immutable file per H000-H020 treatment
  experiments/   staged E00-E16 experiment manifests
  models/        fixed Qwen/LM Studio runtime policy
  embeddings/    independently pinned dense-retrieval model
  scenarios/     robustness perturbations
  backends/      dense-index systems profiles
src/agent_harness/  schemas, LM Studio client, interfaces, metrics, telemetry
tasks/              task manifests and frozen split lists
results/            immutable local raw artifacts, analyses, and reports
docs/               protocol, catalog, metrics, runtime, and paper plan
tests/              standard-library unit tests
paper/              LaTeX manuscript and bibliography
output/pdf/         stable rendered research paper
```

Dense retrieval uses the separately pinned **Qwen3 Embedding 0.6B** model
(`text-embedding-qwen3-embedding-0.6b`) through LM Studio. The local GGUF Q8_0
model produces 1,024-dimensional normalized vectors and exposes a 32,768-token
maximum context. The generative Qwen agent and embedding model remain separate
experimental components.

## Quick start

Create a clean Python 3.11 environment and install the harness plus its pinned
research dependencies:

```bash
python3.11 -m venv .venv
.venv/bin/python -m pip install --upgrade pip
.venv/bin/python -m pip install -r requirements.lock
.venv/bin/python -m pip install -e . --no-deps
```

For the saved-evidence reproduction path, LM Studio and model weights are not
required. See [`REPRODUCING.md`](REPRODUCING.md) for the exact artifact and
inference workflows.

```bash
PYTHONPATH=src python3 -m agent_harness.cli validate
PYTHONPATH=src python3 -m agent_harness.cli list-harnesses
PYTHONPATH=src python3 -m agent_harness.cli plan --experiment E03
PYTHONPATH=src python3 -m agent_harness.cli probe-model
PYTHONPATH=src python3 -m agent_harness.cli probe-embedding --infer
PYTHONPATH=src python3 -m agent_harness.cli run-pilot --experiment E00
PYTHONPATH=src .venv/bin/python -m agent_harness.cli run-live-agent --experiment E07
PYTHONPATH=src .venv/bin/python scripts/preflight_study2.py
PYTHONPATH=src .venv/bin/python -m agent_harness.cli run-study2
PYTHONPATH=src .venv/bin/python -m agent_harness.cli run-study2-reliability
PYTHONPATH=src .venv/bin/python scripts/analyze_confirmatory.py
MPLCONFIGDIR=/tmp/agent-harness-mpl PYTHONPATH=src .venv/bin/python scripts/analyze_live_agent.py
MPLCONFIGDIR=/tmp/agent-harness-mpl PYTHONPATH=src .venv/bin/python scripts/analyze_study2.py
MPLCONFIGDIR=/tmp/agent-harness-mpl PYTHONPATH=src .venv/bin/python scripts/analyze_study3.py
MPLCONFIGDIR=/tmp/agent-harness-mpl PYTHONPATH=src .venv/bin/python scripts/analyze_study4.py
PYTHONPATH=src .venv/bin/python scripts/analyze_study4_ancillary.py
MPLBACKEND=Agg MPLCONFIGDIR=/tmp/agent-harness-mpl PYTHONPATH=src .venv/bin/python scripts/analyze_study5.py
PYTHONPATH=src .venv/bin/python scripts/check_study5_paper_consistency.py
PYTHONPATH=src .venv/bin/python -m pytest -q
```

`probe-model` discovers and validates the exact model key, MLX 4-bit variant,
262,144-token loaded context, and reasoning default. Add `--infer` to perform a
small completion and verify the response semantics after discovery.

## Scientific protocol

Start with the [scientific protocol](docs/SCIENTIFIC_PROTOCOL.md),
[harness catalog](docs/HARNESS_CATALOG.md), [LM Studio policy](docs/LM_STUDIO.md),
[metric and artifact schema](docs/METRICS_AND_ARTIFACTS.md), observed
[runtime baseline](docs/RUNTIME_BASELINE.md), [pilot dataset](docs/PILOT_DATASET.md),
[pilot results](docs/PILOT_RESULTS.md), [memory-safety protocol](docs/MEMORY_SAFETY.md),
and [paper plan](docs/PAPER_PLAN.md).
Study 2 is governed by its [preregistration](docs/STUDY2_PREREGISTRATION.md),
[implementation contract](docs/STUDY2_IMPLEMENTATION.md),
[design audit](docs/STUDY2_DESIGN_AUDIT.json), and
[preflight procedure](docs/STUDY2_PREFLIGHT.md).
Study 3 and 4 use [Study 3 preregistration](docs/STUDY3_PREREGISTRATION.md),
[Study 4 preregistration](docs/STUDY4_PREREGISTRATION.md), and their frozen
design audits, compatibility gate, preflights, and ancillary manifest audit.
Study 5 uses its [prospective preregistration](docs/STUDY5_PREREGISTRATION.md),
frozen cell manifests, and sequential E16 selection ledger.
Configuration files are part of the experimental treatment and become immutable
once a recorded run uses them. Any treatment change receives a new harness ID
or experiment version.

The completed manuscript is
[`output/pdf/dissecting_repository_scale_code_agent_harnesses.pdf`](output/pdf/dissecting_repository_scale_code_agent_harnesses.pdf).
Its source and build instructions are under [`paper/`](paper/). The immutable
local analysis used by the paper is
`results/derived/confirmatory_analysis_174ce71bcbce` plus
`results/derived/e07`, `results/derived/study2`, `results/derived/study3`,
`results/derived/study4`, `results/derived/study4_ancillary`, and
`results/derived/study5`.

## Citation and release

Citation metadata is provided in [`CITATION.cff`](CITATION.cff), and the exact
Zenodo record fields are in [`.zenodo.json`](.zenodo.json). The canonical
archival DOI is
[10.5281/zenodo.21781711](https://doi.org/10.5281/zenodo.21781711).
Publication upload instructions are in
[`README_ZENODO.md`](README_ZENODO.md).

The repository intentionally contains one canonical PDF. Raw trajectories,
repository checkouts, index caches, and model weights are excluded from Git and
from the compact release; raw trajectories belong in a separate
checksum-addressed Zenodo archive.

## Licensing

Original software is MIT licensed under [`LICENSE`](LICENSE). The manuscript,
documentation, figures, tables, and original derived research artifacts are CC
BY 4.0 under [`LICENSE-DATA`](LICENSE-DATA). Benchmark patches retain their
upstream licenses; see [`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md).