Google boosts SWE-bench performance via harness optimization
Google researchers have introduced a novel approach to improving AI agent performance, bypassing traditional fine-tuning for a focus on recursive self-improvement of agent harnesses. By allowing an optimizer to iteratively rewrite the agent's operating harness—including system prompts, memory, and tools—around a frozen base model like Gemini 3.8 Flash, the system demonstrates significant gains on coding benchmarks like SWE-bench Lite. This contrastive, execution-guided optimization enables agents to dynamically adapt to complex tasks without human intervention, setting a new paradigm for efficient, high-performance agent engineering.
This is a shift in the agentic paradigm from "model-centric" to "harness-centric".
- –**Performance vs. Cost:** By keeping the base model (Gemini 3.8 Flash) frozen and focusing computational overhead on optimizing the harness, Google achieves high performance with lower cost and latency than fine-tuning or scaling up models.
- –**Scalability:** The *Propose-Probe-Compose* loop allows for automated, recursive improvements, potentially enabling agents to learn and adapt to new coding environments or tasks autonomously.
- –**Benchmarking Insight:** Demonstrating consistent gains on SWE-bench Lite suggests that harness optimization may be a more accessible pathway to state-of-the-art performance for developers than model training.
- –**Future-proofing:** The use of rationales in the memory layer (SIM) hints at future agents that don't just "act" but "reason" about their own structural improvements, a critical step toward fully autonomous software engineers.
DISCOVERED
4h ago
2026-11-18
PUBLISHED
4h ago
2026-11-18
RELEVANCE
AUTHOR
omarsar0