EvoAgentBench

01 — OVERVIEW

EvoAgentBench

Benchmarking Agent Self-Evolution via Ability Transfer

EvoAgentBench evaluates whether agents can transfer reusable procedures from past experience to new tasks. Its Ability-guided split covers web research, algorithmic reasoning, software engineering, and knowledge work, with verified training-side support for every test task.

🌐

Multi-Domain Evaluation

4 agentic domains — web research, algorithmic reasoning, software engineering, and knowledge work — with 528 training tasks and 267 supported test tasks.

🤖

Multi-Agent Support

Plug in any CLI-based agent — Nanobot, OpenClaw, or your own. Each task runs in isolated config with independent workspace, supporting concurrent execution and automatic retry.

🧬

Self-Evolution Comparison

A standardized train → evolve → evaluate protocol comparing Memento, ReasoningBank, GEPA, and the diagnostic Anchor Skill† reference under matched conditions.

02 — LEADERBOARD

Method ranking per configuration

For each (agent, model, domain) cell, methods are sorted by Δ gain (method score − Vanilla). The Vanilla score is shown next to each method.

Agent
Model
Domain
positive Δnegative Δ24 configurations

Web Research

OpenClaw · 27B
method
van.
Δ
GEPA
10.8
+4
Anchor Skill†
10.8
+4
ReasoningBank
10.8
+2
Memento
10.8
+1

Algorithmic Reasoning

OpenClaw · 27B
method
van.
Δ
Anchor Skill†
51.9
+9
ReasoningBank
51.9
+6
Memento
51.9
+2
GEPA
51.9
+2

Software Engineering

OpenClaw · 27B
method
van.
Δ
Anchor Skill†
42.3
+10
Memento
42.3
+3
GEPA
42.3
-1
ReasoningBank
42.3
-5

Knowledge Work

OpenClaw · 27B
method
van.
Δ
Anchor Skill†
43.6
+5
ReasoningBank
43.6
+5
Memento
43.6
+1
GEPA
43.6
-12

Web Research

OpenClaw · 397B
method
van.
Δ
Anchor Skill†
31.8
+13
ReasoningBank
31.8
+7
Memento
31.8
+2
GEPA
31.8

Algorithmic Reasoning

OpenClaw · 397B
method
van.
Δ
Anchor Skill†
52.7
+11
GEPA
52.7
+9
Memento
52.7
+8
ReasoningBank
52.7
+7

Software Engineering

OpenClaw · 397B
method
van.
Δ
Anchor Skill†
66.7
+10
GEPA
66.7
+1
Memento
66.7
-3
ReasoningBank
66.7
-5

Knowledge Work

OpenClaw · 397B
method
van.
Δ
Anchor Skill†
55.8
+4
ReasoningBank
55.8
+2
GEPA
55.8
+1
Memento
55.8
-6

Web Research

OpenClaw · Gemma-4-31B
method
van.
Δ
Anchor Skill†
13.3
+5
GEPA
13.3
+4
Memento
13.3
+3
ReasoningBank
13.3

Algorithmic Reasoning

OpenClaw · Gemma-4-31B
method
van.
Δ
Anchor Skill†
72.1
GEPA
72.1
-2
Memento
72.1
-4
ReasoningBank
72.1
-11

Software Engineering

OpenClaw · Gemma-4-31B
method
van.
Δ
GEPA
25.0
+17
Anchor Skill†
25.0
+8
ReasoningBank
25.0
+7
Memento
25.0
-5

Knowledge Work

OpenClaw · Gemma-4-31B
method
van.
Δ
Anchor Skill†
44.1
+11
GEPA
44.1
+1
ReasoningBank
44.1
-1
Memento
44.1
-11

Web Research

Nanobot · 27B
method
van.
Δ
Anchor Skill†
8.7
+5
Memento
8.7
+4
GEPA
8.7
+3
ReasoningBank
8.7
+1

Algorithmic Reasoning

Nanobot · 27B
method
van.
Δ
Anchor Skill†
52.7
+11
ReasoningBank
52.7
+10
GEPA
52.7
+8
Memento
52.7
+4

Software Engineering

Nanobot · 27B
method
van.
Δ
ReasoningBank
45.8
+7
Anchor Skill†
45.8
+4
GEPA
45.8
+3
Memento
45.8
-36

Knowledge Work

Nanobot · 27B
method
van.
Δ
Anchor Skill†
43.9
+14
ReasoningBank
43.9
+5
Memento
43.9
+4
GEPA
43.9
+3

Web Research

Nanobot · 397B
method
van.
Δ
Anchor Skill†
12.8
+7
Memento
12.8
+4
ReasoningBank
12.8
+3
GEPA
12.8
+2

Algorithmic Reasoning

Nanobot · 397B
method
van.
Δ
Anchor Skill†
57.8
+12
GEPA
57.8
+7
ReasoningBank
57.8
+7
Memento
57.8
+6

Software Engineering

Nanobot · 397B
method
van.
Δ
Anchor Skill†
62.5
+17
GEPA
62.5
+9
Memento
62.5
+2
ReasoningBank
62.5
+1

Knowledge Work

Nanobot · 397B
method
van.
Δ
Anchor Skill†
53.4
+9
ReasoningBank
53.4
-1
Memento
53.4
-1
GEPA
53.4
-1

Web Research

Nanobot · Gemma-4-31B
method
van.
Δ
Anchor Skill†
9.2
+3
Memento
9.2
+1
GEPA
9.2
+1
ReasoningBank
9.2
-5

Algorithmic Reasoning

Nanobot · Gemma-4-31B
method
van.
Δ
Anchor Skill†
69.0
+4
ReasoningBank
69.0
+4
Memento
69.0
+3
GEPA
69.0
+1

Software Engineering

Nanobot · Gemma-4-31B
method
van.
Δ
GEPA
17.3
+20
Anchor Skill†
17.3
+10
ReasoningBank
17.3
+7
Memento
17.3
+4

Knowledge Work

Nanobot · Gemma-4-31B
method
van.
Δ
Anchor Skill†
43.2
+6
GEPA
43.2
+5
ReasoningBank
43.2
+3
Memento
43.2
+3

Bar length encodes Δ magnitude. See the full leaderboard for per-cell scores and the paper’s turn-cost analysis.

03 — WHY THIS MATTERS

Why Self-Evolution Matters, and What We Learned Running This

Why self-evolution matters

If an agent solves a problem today, it shouldn’t have to start from zero on a similar one tomorrow. Useful experience isn’t just a log of what happened — it’s a way of working: a search habit, a debugging move, a verification step, a recipe for producing something useful. Self-evolution is the question of whether an agent can pick up these habits on its own, from its own past attempts, without retraining the model underneath.

Most benchmarks today don’t really test this. They either ask “can the agent solve a fresh task?” or “can the agent remember what it saw?” Neither tells you whether yesterday’s way of working actually shows up when the agent tries something new today. EvoAgentBench is built around that specific question.

What the experiments show

Across two agent scaffolds, three backbones, and four domains, the results expose where current self-evolution systems still break.

Ability content transfers across model families. Anchor Skill† improves every scaffold–backbone–domain cell even though its construction backbones are disjoint from the evaluation backbones. The procedural content is transferable when it is extracted and routed correctly.

Every automatic method still shows negative transfer. Memento, ReasoningBank, and GEPA each improve some settings, but none remains positive in every cell. One striking mismatch is Memento on Nanobot / Qwen3.5-27B / software engineering, where performance drops by 36.3 points.

Extraction and routing remain the bottleneck. Automatic methods see the same training prompts, trajectories, and verifier outcomes. Their uneven gains therefore point to how reusable content is extracted, indexed, and delivered at test time, not to missing training-side support.

What we did to make the comparisons fair

For numbers to be worth comparing, the setup has to be tight. Three things we cared about most:

Nobody gets to peek at the answers. Methods can only look at training tasks: the question, what the agent tried, and whether it worked. The actual test answers, and the test trajectories that succeeded, are off-limits during evolution. We didn’t make this a rule for people to follow; we wired it in so the two paths simply don’t meet.

Training and test tasks are related, not random. The Ability Graph links tasks through trace-grounded procedural overlap. Every test task has verified support from training-side Abilities, so a failed transfer cannot be dismissed as an unsupported test case.

Every cell on the leaderboard is real. Two agent frameworks, three backbones, four domains, and four evolution conditions produce 96 method comparisons, all measured against matched Vanilla baselines. Tasks, tools, timeouts, scoring, and base agent configuration are fixed across methods; results are averaged over three independent runs per instance.

04 — DOMAINS

Evaluation Domains

EvoAgentBench builds an Ability Graph over four agentic domains and selects a 528/267 train/test split with verified training-side support for every test task.

DomainBase BenchmarkDescriptionAbilitiesCommunitiesAbilities / TaskTrainTest
Web ResearchBrowseComp-PlusSearch a controlled corpus and verify evidence for multi-constraint questions.38132.4215465
Algorithmic ReasoningLiveCodeBenchSolve programming problems against hidden test cases with pass@1 scoring.58222.1218286
Software EngineeringSWE-Bench VerifiedRepair real-world repositories and pass the designated hidden test suites.38152.218756
Knowledge WorkGDPValProduce professional deliverables evaluated against expert references.3662.6310560

05 — SELF-EVOLUTION

Evolution Conditions

EvoAgentBench provides a standardized protocol for evaluating agent self-evolution methods — techniques that let agents learn from past experience and improve future performance. The paper evaluates three automatic methods plus Anchor Skill†, a diagnostic reference.

Memento

Retrieval-based (CBR)

Retrieves the closest retained training case, using surface similarity to route prior task experience.

GitHub

ReasoningBank

Memory + reasoning

Distills natural-language strategies from training trajectories and retrieves them by strategy similarity.

GitHub

GEPA

Prompt evolution

Evolves one prompt on the training split and broadcasts it to every test task without per-task retrieval.

GitHub

Anchor Skill†

Diagnostic reference

Diagnostic reference using train-side Ability-grounded skill content and curator-side Ability-label routing. It is not a deployable automatic method.