RL Environments and the Hierarchy of Agentic Capabilities
Our RL environment run on 9 models revealed the core capabilities all agents need to master: tool use, planning, adaptability, groundedness, and common sense.
Long-Context Agentic Instruction Following
HANDBOOK.md Agents
Can an agent follow a 100-page company handbook inside an enterprise RL environment?
HANDBOOK.md is a benchmark for long-context agentic instruction following, modeled on how professionals follow corporate policy in their day-to-day work. Each task is a unique RL environment with internal tools and external MCP servers, across five enterprise domains.
HANDBOOK.md Agents Leaderboard
1
Claude
Fable 5.1 (Adaptive/Max)
38.8
%
1
Claude
Fable 5 (Adaptive/Max)
36.2
%
1
Claude
Opus 5 (Adaptive/Max)
32.3
%
1
Grok
4.6 (xHigh reasoning)
27.3
%
1
DeepSeek
V4 Pro (Max reasoning)
26.9
%
1
GPT
5.6 Sol (Max reasoning)
23.5
%
1
DeepSeek
V4 Flash Vision (experimental) (Max reasoning)
23.1
%
1
Claude
Opus 4.8 (Adaptive/Max)
21.9
%
1
GPT
5.5 (xHigh reasoning)
21.5
%
1
Qwen
3.8 Max (xHigh reasoning)
16.5
%
1
Grok
4.5 (High reasoning)
15.8
%
1
Muse
Spark 1.3 (xHigh reasoning)
15.4
%
1
GLM
5.3 (Max reasoning)
15.0
%
1
Muse
Spark 1.1 (xHigh reasoning)
13.5
%
1
Muse
Spark 1.2 (xHigh reasoning)
13.1
%
1
Kimi
K3 (Max reasoning)
11.9
%
1
Gemini
3.5 Flash (High reasoning)
11.2
%
1
Gemini
3.7 Flash (High reasoning)
10.8
%
1
Gemini
3.8 Flash (High reasoning)
10.8
%
1
Claude
Sonnet 4.6 (Adaptive/Max)
10.4
%
1
GLM
5.3 Flash (Max reasoning)
10.4
%
1
GLM
5.2 (xHigh reasoning)
10.0
%
1
Gemini
3.1 Pro (High reasoning)
10.0
%
1
DeepSeek
V4 Pro (preview) (xHigh reasoning)
9.2
%
1
Qwen
3.7 Max (Thinking on)
8.5
%
1
Hy
Hy3 (High reasoning)
7.7
%
1
DeepSeek
V4 Flash (preview) (xHigh reasoning)
7.3
%
1
DeepSeek
V4 Flash (xHigh reasoning)
7.3
%
1
Kimi
K2.6 (Thinking on)
6.9
%
1
Gemini
3.6 Flash (High reasoning)
5.0
%
1
Muse
Glimmer 30B (xHigh reasoning)
3.5
%
1
Gemini
3.5 Flash-Lite (High reasoning)
3.1
%
1
Inkling
Inkling (xHigh reasoning)
2.3
%
1
Grok
4.3 (High reasoning)
1.9
%
1
Nemotron
3 Ultra
1.5
%
1
Inkling
Small (xHigh reasoning)
1.2
%
1
Nemotron
Lightning 3.5 30B A3B (Thinking on)
0.0
%
VIEW ALL BENCHMARKS
Agentic Coding
sudo L7
Coding agents are getting very good at being L3s. But do they have the judgment to become L7s? Across 60 tasks, agents must navigate unfamiliar systems, choose architectures, spot hidden risks, and know when technically correct code is still the wrong thing to ship.
Rank
Model
Score
41.7
%
Claude Sonnet 5.5 (Adaptive/Max)
Anthropic
41.7
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
33.9
%
GPT 6 Astra (Max)
OpenAI
31.1
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
28.9
%
GPT 6.1 Sol (Max)
OpenAI
20.6
%
Qwen 3.8 Max
Alibaba Cloud
Agentic Finance
DAYJOB: Finance
DAYJOB: Finance measures long-horizon finance agents across corporate finance, banking, credit, investing, and real assets. It's built around the messy reality of knowledge work: realistic requests, complex environments, and end-to-end professional judgment.
Rank
Model
Score
23.9
%
Claude Opus 5.5 (Adaptive/Max)
Anthropic
21.5
%
GPT 6 Astra (Max)
OpenAI
20.3
%
Gemini 4 Argon (High)
Google
19.8
%
Claude Fable 5.1 (Adaptive/Max)
Anthropic
14.8
%
Muse Spark 1.3 (Max)
Meta
14.5
%
Grok 4.7 (xHigh)
xAI