Loading…
Six Agent Harness Capabilities for Higher Model Performance
Michelle Horton
- Source
- NVIDIA
- Published
- Added to Yomu
Summary
NVIDIA Labs released NVIDIA Labs Object-Oriented Agents (NOOA), an open-source research preview treating AI agents as single Python classes. The harness implements six model-facing capabilities: typed inputs and outputs, passing live objects by reference, code-based actions, programmable loops, explicit object state, and model-callable harness APIs. Instead of context compaction, passing live references keeps transcripts append-only, preserving prompt cache validity across long sessions. NOOA also includes a curated SQLite-backed long-term memory system forming a typed knowledge graph. Across evaluations, NOOA achieved 82.2 percent on SWE-bench Verified with GPT-5.5 using roughly 1.1M tokens per task, 86.8 percent on CyberGym L1 without network access, and 85.1 percent mean RHAE on ARC-AGI-3 using GPT-5.6-sol at under twenty dollars per game.
Context
Harness design shapes agent outcomes, token efficiency, and benchmark scores as much as underlying models, but existing agent development requires coordinating disjoint prompt templates, tool schemas, callback code, and workflow graphs.
Approach / What changed
NVIDIA Labs created NOOA, an open-source Python framework where an agent is a single class with docstrings as prompts, typed methods as capabilities, and ellipsis-bodied methods completed at runtime by an LLM loop. It uses pass-by-reference execution to avoid context serialization, alongside a model-curated SQLite knowledge graph memory system.
Takeaways
- Passing live Python objects by reference rather than serializing data into context keeps transcripts append-only and eliminates the need for context compaction passes on SWE-bench Verified.
- NOOA reached 82.2% on SWE-bench Verified with GPT-5.5 using 29 LLM calls and ~1.1M tokens per task, cutting token costs roughly in half compared to 66-call harnesses scoring 78.2%.
- NOOA's curated long-term memory persists as an inspectable SQLite knowledge graph with typed relationships such as supports, contradicts, and derived-from.