Google has just open-sourced Harness2. It turns any AI harness (scaffolding/runtime framework) into a system that improves itself over time.
A “harness” refers to everything built around a model: system prompts, tools, memory, skills, and subagents. Typically, these are designed once by humans, and the agent continues to reuse this fixed configuration across every new task.
With Harness2, however, the agent optimizes its own harness on a per-task basis. It proposes improvements, runs multiple candidate versions, and selects the better option by comparing their actual performance. The entire process requires neither ground truth answers nor graders, and the model weights remain completely untouched.
What the agent learns on one task transfers to the next. Once a task is completed, it saves the improved harness along with valuable takeaways, giving subsequent optimizations a head start rather than starting from scratch.
Google evaluated this across legal, office, and professional work benchmarks using Gemini 3.5 Flash and GLM-5. The results show that even a single round of adjustments outperformed the original harness, while further optimization rounds pushed scores even higher.
GitHub: https://github.com/google-research/harness2
Top Comments
CoolConfusion434 (⬆159): To quote Mr. Krabs: I feel a disturbance deep down in my context window. All this “self-optimization” has to live somewhere, and it’s bound to end up eating into context. I’m sure it’s fine for cloud users sitting on 1M context windows, but those of us running local models are going to get squeezed dry. That said, it’s great to see another open-source harness, especially from a giant like Google.
FaceDeer (⬆39): It alters context, but it doesn’t necessarily bloat it. In fact, I’d argue one of its core jobs is trimming out the redundant or unnecessary bits.
abhi91 (⬆64): Great in theory. Poetiq has already shown that this essentially breaks benchmarks because the harness can end up overfitting to specific tasks. That said, for repetitive workloads like customer support, enterprises could put this to great use right now.
acoolglassofwater (⬆26): Can an actual LLM expert (no armchair speculation, please) weigh in on what this actually means? I understand the basics, but sentiment around “self-improving agents” is always so polarized here. Is this release just a hyped-up nothingburger, or does it genuinely represent an interesting and meaningful step forward?
itsmebenji69 (⬆30): I’d call it moderately interesting. A harness definitely places an upper bound on an agent’s performance on a task, so continuously modifying the harness to better fit the task is valuable—it breaks that ceiling (if the harness was initially holding back performance, a few rounds of iteration can fix that). But the obvious caveat is that heavily tailoring it to Task A might degrade performance on Task B. It’s essentially overfitting. To give a simple example: Task A is calculating x + y, and Task B is calculating x * y. The first version of the harness only provides x and y, and…
Holiday_Point_603 (⬆28): Pretty funny that Google is demoing their new software on six-month-old models, haha.
Spara-Extreme (⬆37): Google is massive. The team working on this project is completely separate from the folks building Gemini 4.
DifficultIndustry148 (⬆42): They’re just showing that the tool works with models outside their own ecosystem. No need to read too much into it.