Harness Optimizers and Meta-Harnesses

Harness optimizers are adjacent to the strict runtime map. They do not primarily execute user tasks or coordinate a fleet. They change a target harness, evaluate candidates, and decide which version should govern later runs. Most entries below are research systems, not production products.

System	Optimizes	Evidence	Maturity and main limit
ADAS	Code-defined prompts, tools, and workflows	ICLR 2025 experiments across coding, science, and math	Peer-reviewed, but headline comparisons are system-level rather than pure fixed-model ablations
AFlow	Workflow topology represented as code	5.7% average improvement across six benchmarks	ICLR 2025; heterogeneous baselines
ACE	Context and memory playbooks	+10.6% on agents and +8.6% on finance tasks	ICLR 2026; narrower than whole-harness optimization
GEPA	Prompts using reflected trajectories	6% average gain over GRPO across six tasks; up to 35 times fewer rollouts	ICLR 2026 Oral; prompt-level optimizer
Meta-Harness	End-to-end harness code	Classification, math, and TerminalBench-2 improvements	2026 preprint; coding search and final evaluation reuse the same 89 tasks
Agentic Harness Engineering	Prompt, tools, middleware, skills, subagents, and memory	Terminal-Bench 2 pass@1 from 69.7% to 77.0% over ten iterations	2026 preprint focused on coding benchmarks
HarnessOpt-Bench	Evaluates the optimizer, not one target harness	Five optimizer models, four downstream tasks, 111 scored runs	2026 benchmark preprint; early protocol awaiting broader reproduction

Do not fold these systems into the 42-runtime count. A runtime owns the task loop. An optimizer owns a search loop over candidate harnesses. Agent Harness Engineering §11 documents the evidence and the minimum evaluation protocol.