Google Paper: Agent Harness Memory Lifts Score 73.4 to 84.8

Researchers at Google Cloud AI and Google DeepMind report that giving an AI agent’s harness optimizer a memory of past edits lifted a workspace score to 84.75, against 73.42 without it, after the underlying model was swapped from Claude Opus 5.5 to Gemini 3.8 Flash, in a preprint posted to arXiv on 13 November 2026.

Our number: that is an 11.33-point gap, about 15% above the no-memory result (our arithmetic), from the same 68.41 starting score.

Business Pill · Least Privilege

An evergreen Business Engineer explainer on access control for software agents. It teaches the general idea only and does not discuss this paper.

The key insight: A Google preprint reports that recording why harness edits were made let an agent keep improving after a model swap: 84.75 with the memory against 73.42 without, an 11.33-point gap, and higher scores in 9 of 10 coding runs.

What the Paper Proposes

An agent’s harness is the code around a model that manages its context, tools and execution. The paper studies what it calls “Continual Recursive Self-Improvement (CRSI)”: continuing an automated harness-optimization run when the benchmark or the model changes.

Its method is “Self-Improvement Memory (SIM) for the harness optimizer, rather than the task-solving policy”. SIM “groups edits by recurring harness problem and records their rationale, predicted and observed outcomes, and evaluation context”, and “New attempts extend this history as optimization continues”.

In the preprint's workspace test, a harness evolved on Claude Opus 5.5 and continued on Gemini 3.8 Flash reach
In the preprint’s workspace test, a harness evolved on Claude Opus 5.5 and continued on Gemini 3.8 Flash reached 84.75 with Self-Improvement Memory and 73.42 without it, from the same 68.41 start. Source: arXiv 2611.18856.

What Was Posted on X

kimmonismus posted on 18 November 2026 about the Google preprint, summarising that with the log of changes the agent scored 84.75 on workplace tasks against 73.42 without it after Opus 5.5 was swapped for Gemini 3.8 Flash, and came out ahead in 9 of 10 coding runs.

What the Results Show

In the model-swap test, a harness evolved with Claude Opus 5.5 was continued on Gemini 3.8 Flash. The paper says: “When the Opus-evolved harness is first run with Gemini, it scores below Gemini’s seed harness in every domain”, with deficits from 0.60 to 5.48 points.

After continuing, “Full SIM finishes highest in all four and exceeds harness-only continuation by 3.60–11.33 points”. The largest gap was in the agentic workspace domain: “both begin at 68.41, but Full SIM reaches 84.75 while Harness Transfer reaches 73.42”.

In coding, across ten independent source runs, “SIM finishes higher in 9 of 10 pairs”, with a mean advantage of 2.36 points (paired p=0.003, as the paper reports).

How It Fits This Week’s Self-Improvement News

The approach differs from Meta’s self-taught memory, which moved gains into model weights: here the memory sits with the optimizer, so a new model can inherit why earlier harness edits were made. It lands the same week as OpenAI’s first ‘High’ self-improvement rating, remarks on self-improving AI by OpenAI’s Chen and Schulman and Microsoft AI chief Mustafa Suleyman calling AI improving AI dangerous.

The Structural Read

As we read the paper, the asset that carries over between model generations is the record of what was tried and why, not the harness code alone.

As we read it, that makes harness work cumulative: a lab that swaps models does not have to restart optimization from its seed harness.

The preprint’s authors, arXiv 2611.18856

“both begin at 68.41, but Full SIM reaches 84.75 while Harness Transfer reaches 73.42”

Three Implications

FOR AGENT BUILDERS The paper reports that an Opus-evolved harness first scored below Gemini’s seed harness in every domain when moved across models.

FOR SELF-IMPROVEMENT The paper reports SIM finished highest in all four domains in the cross-model comparisons.

FOR EVALUATION The results are the authors’ own, on their benchmarks, and not independently replicated.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The analysis puts it this way: “Foundation models become relatively commoditized as orchestration layers capture differentiation.”

As we read the preprint, a harness memory that survives a model swap is one mechanism by which that layer could hold its value. The analysis does not discuss this paper.

What Is Not Established

These are the authors’ own results on their chosen benchmarks; the paper is a preprint and we have not seen independent replication or peer review. The paper notes the work was done while its lead author was an intern at Google; it is a research paper, not a product announcement, and we did not see Google say it uses the method in a product.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Agentic Harness War

A Business Engineer analysis of how agent harnesses and orchestration layers compete as models commoditize.

Read the Map of AI →

The Bottom Line

A Google preprint reports that remembering why harness edits were made let an agent keep improving after a model swap, reaching 84.75 against 73.42 without the memory, an 11.33-point gap.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. All results and quotations are from the arXiv preprint 2611.18856 (13 November 2026), read by us; the authors are affiliated with Google Cloud AI, Google DeepMind, Google Cloud AI Research and Penn State. The 11.33-point gap is the paper’s own range end; the about 15% relative gain is our arithmetic. We did not contact Google.

Sources: arXiv 2611.18856, Self-Improvement Memory Enables Continual Recursive Self-Improvement (13 Nov 2026) · kimmonismus on X, 18 Nov 2026

Scroll to Top
FourWeekMBA