Improving an agent does not always require changing its model weights. The Meta-squared paper studies the harness around a frozen model: the prompts, tools, memory and hooks that shape how a task is carried out.
Learning guidance rather than one shared harness
Offline optimisation can produce a single harness that works well on training tasks but adapts poorly to a new run. Online evolution can respond to new evidence, yet may have to decide how to intervene without a reliable verifier.
Meta-squared learns a reusable textual Meta-Skill from contrasting training trajectories. Those contrasts include runs without an evolver, runs using the current Meta-Skill and runs given privileged verifier information that is available only during training.
Keep the harness or change it?
During online use, an evolver combines the frozen Meta-Skill with evidence from the current trajectory. It chooses whether to keep the harness or make a task-local edit. An edit affects the rest of that run without resetting the environment or the progress already made.
The reusable guidance transfers across tasks, while the edits remain local to the task being attempted. Held-out verifier feedback is not available to the evolver at this stage.
What the benchmarks show
With GPT-5.6-Terra, the paper reports 81.00% on Terminal-Bench 2.1 and 44.44% on DeepSWE 1.1, the highest results among the methods evaluated in that study. Ablations support both contrastive guidance and keeping edits task-local.
Where the distinction matters
This is a method for adjusting an agent's operating setup, not evidence that its underlying model recursively trains a successor. For developers investigating the approach, a useful evaluation should track when edits help, when leaving the harness unchanged is better and whether changes preserve the original task requirements. Benchmark gains do not remove the need to inspect those decisions on new workloads.
Comments