To be clear: this post is not telling you to install anything. On 9 October 2026, Google open-sourced a research framework called Harness². The point is not another stronger model — it is a sentence worth remembering: an agent can get better without a better model, by changing the settings around it.
What is a harness? One analogy
Imagine you hire a craftsman. Their skill is the model. How well the job goes also depends on everything else: how you brief them, which tools they get, whether they remember what you said last week, what they are not allowed to decide alone, and when they must come back and ask.
All of that — everything except the craftsman — is the harness. The label does not matter. What matters: the harness is the shell that decides behaviour; the model is only the engine.
The same engine with different gearing and chassis performs very differently. That is what Harness² sets out to demonstrate.
It edits eight files
System prompt (role, tone, boundaries)
Guardrails (what it must not do)
Memory (what carries across conversations)
Tool guidance (when to use which tool)
Subagents (delegation)
Skills (reusable procedures)
Scripts
Plugins (middleware)
The model weights stay frozen. Not a single parameter is trained. Only those eight files change.
How does it know an edit helped?
This is the practical heart of the paper — a loop called Propose–Probe–Compose:
Propose: two candidate edits, one aimed at a general weakness, one at the task at hand
Probe: both are actually run on the same task and compared against the original run
Compose: keep only what the observed behaviour supports; discard the rest
The key: no ground-truth answer key is required, which is what blocks self-improvement on real, open-ended work. The judge is whether behaviour improved, not whether the answer was right.
There is a second layer: after each task the lessons go into a library that guides future edits. Remove that library and the score falls from 77.5 to 74.4 — the accumulation of experience is doing real work.
The published numbers (not independently tested by us)
GLM-5 with OpenCode: 64.2 → 75.7 on WorkBuddy-Bench (+11.5)
Same setup on the LAB benchmark: 59.6 → 71.4 (+11.8 points)
All four model–harness configurations improved on all three benchmarks
Cost: about four agent runs per task (reference, two probes, final)
Double-digit gains without touching the model. But four times the compute — that is not a free upgrade.
What it means for you
One: the model is only part of the answer
The first question most buyers ask about an AI assistant is which model it uses. Fair, but insufficient. With the same model, the quality of guardrails, the cleanliness of the knowledge base, memory, and handover rules often matter more than the model version.
Two: stop chasing model releases
If your process, boundaries and knowledge base are a mess, upgrading the model repeatedly will not fix it. A modest model with clear instructions is usually steadier than a strong model with contradictory ones.
Three: ask three questions when buying
How do you define what the bot must not answer?
Are conversations logged, for how long, and who can read them?
What always goes to a human?
Suppliers who answer these clearly have usually shipped something real.
Why not put it in production
To be blunt, this is research code, not a product:
Version 0.1.0; the README states it is not an officially supported Google product
Bound to Vertex AI; needs Linux with Python 3.12/3.13 and uv
The README warns that executor network access is unrestricted — isolate it before running
Some benchmarks improve only marginally (JobBench with Codex: about +1 point)
So the right use is to borrow the idea, not the code: treat the shell around the model as something you version, test and improve — the way you improve website copy.
Our view
The direction matches how we work. AI projects succeed or fail on whether someone wrote the rules down, someone reviews the output, and problems can be traced — not on the model. That is why we spend our time agreeing boundaries with clients rather than advertising the newest model.
If you want a second opinion on whether your current assistant, or the one you are planning, is set up properly, book a 30-minute call. Prices are public on the pricing page.