---
type: "post"
id: "cline-agents-harness-unlocks-benchmark-leading-performance-from-03fd56"
slug: "cline-agents-harness-unlocks-benchmark-leading-performance-from-03fd56"
title: "Cline Agents Harness Unlocks Benchmark-Leading Performance from NVIDIA Nemotron 3.5 Super VL"
description: "Cline reports that Nemotron-3.5 Super VL 120B A12B scored 61.8% on Terminal-Bench 2.1 with Cline’s harness and thinking mode off, versus 53.9%, 51.7%, and 44.9% in three other open-source harnesses."
retrieval_nugget: "Cline reports that Nemotron-3.5 Super VL 120B A12B scored 61.8% on Terminal-Bench 2.1 with Cline’s harness and thinking mode off, versus 53.9%, 51.7%, and 44.9% in three other open-source harnesses."
published_at: "2026-10-15"
updated_at: "2026-10-15"
record_date: "2026-10-15"
date_kind: "published_at"
topics: ["agent-harness","coding-agents","harness-engineering"]
source_urls: ["https://cline.ghost.io/nvidia-nemotron-3-5-super-vl-cline-agent-harness"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:cline-agents-harness-unlocks-benchmark-leading-performance-from-03fd56"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:50f25daa175dcf8e412ec8a4452af2b66d9e1a4f54ecd5b73dfc2e5adff283f1","reviewed_at":"2026-10-15T16:09:22.316Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":2,"watch_condition_count":2,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"harness design can materially shape observed agent performance even when the underlying model is held constant.","observed_facts":[{"text":"Cline reports 61.8% for Nemotron-3.5 Super VL 120B A12B on Terminal-Bench 2.1 with Cline and thinking mode off.","source_urls":["https://cline.ghost.io/nvidia-nemotron-3-5-super-vl-cline-agent-harness"]},{"text":"Three other open-source harnesses scored 53.9%, 51.7%, and 44.9%; two used high reasoning or effort settings.","source_urls":["https://cline.ghost.io/nvidia-nemotron-3-5-super-vl-cline-agent-harness"]},{"text":"Cline says its internal Nemotron-based Slack agent triages tasks and investigates user-reported bugs.","source_urls":["https://cline.ghost.io/nvidia-nemotron-3-5-super-vl-cline-agent-harness"]}],"mechanism":"Cline combines a plan-before-code instruction, a stall detector that emits structured events, and checks aimed at the behavior each task requires.","why_now":"open-weight models are increasingly being evaluated as parts of runnable agent systems, so harness comparisons influence how teams judge practical capability.","implications":["Teams comparing agent models should keep the harness and reasoning settings visible rather than treating benchmark scores as properties of the model alone.","Task-level verification examples can clarify what a benchmark score represents, while remaining distinct from independent validation."],"evidence_boundary":"Cline’s post reports its own benchmark runs and internal Slack agent use; it does not provide an independent replication, isolate feature effects, or measure token savings.","watch_conditions":["whether independent evaluations reproduce the gap under matched settings and disclose task-level results.","whether future comparisons isolate planning, stall detection, and verification as separate contributors."],"related_records":[],"new_branch_reason":"No prior New Runtime record was supplied; this source supports a distinct record about a same-model harness comparison and reported operational use."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/cline-agents-harness-unlocks-benchmark-leading-performance-from-03fd56/","markdown":"https://newruntime.com/posts/cline-agents-harness-unlocks-benchmark-leading-performance-from-03fd56.md","json":"https://newruntime.com/posts/cline-agents-harness-unlocks-benchmark-leading-performance-from-03fd56.json"}
related_patterns: ["harness-architecture-outlives-model-choice","ai-engineering-evaluations-are-splitting-into-task-specific"]
---

# Cline Agents Harness Unlocks Benchmark-Leading Performance from NVIDIA Nemotron 3.5 Super VL

## Retrieval answer

Cline reports that Nemotron-3.5 Super VL 120B A12B scored 61.8% on Terminal-Bench 2.1 with Cline’s harness and thinking mode off, versus 53.9%, 51.7%, and 44.9% in three other open-source harnesses.

Cline reports that Nemotron-3.5 Super VL 120B A12B scored 61.8% on Terminal-Bench 2.1 with Cline’s harness and thinking mode off, versus 53.9%, 51.7%, and 44.9% in three other open-source harnesses. This is evidence that harness choice can materially change a model’s measured performance on a coding-agent benchmark, not proof that Cline is generally superior across tasks. The exact comparison is vendor-reported, and two competing runs used high reasoning or effort settings.

The thesis is: harness design can materially shape observed agent performance even when the underlying model is held constant. The mechanism is: Cline combines a plan-before-code instruction, a stall detector that emits structured events, and checks aimed at the behavior each task requires. The company’s examples include a chess task solved by parsing the board into FEN and searching for checkmates, plus headless-terminal checks completed in five commands. Those details make the result more interpretable than a score alone, but they do not establish which component caused the score gap.

Why now: open-weight models are increasingly being evaluated as parts of runnable agent systems, so harness comparisons influence how teams judge practical capability. Teams comparing agent models should keep the harness and reasoning settings visible rather than treating benchmark scores as properties of the model alone. What to watch: whether independent evaluations reproduce the gap under matched settings and disclose task-level results. The evidence boundary is: Cline’s post reports its own benchmark runs and internal Slack agent use; it does not provide an independent replication, isolate feature effects, or measure token savings. Its internal agent triages tasks and investigates user-reported bugs, including a reported Windows diagnostics export failure. That operational example shows use, but not a quantified productivity outcome.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Agent Harness](https://newruntime.com/topics/agent-harness.md): Explore the agent-harness topic hub.
- topic hub: [Coding Agents](https://newruntime.com/topics/coding-agents.md): Explore the coding-agents topic hub.
- topic hub: [Harness Engineering](https://newruntime.com/topics/harness-engineering.md): Explore the harness-engineering topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)