DAYJOB: Law Can AI do a lawyer’s job?

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.

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.
HANDBOOK.md Agents Leaderboard
1
Claude
Opus 5.5 (Adaptive/Max)
52.3
%
1
Claude
Fable 5.2 (Adaptive/Max)
52.3
%
1
Gemini
4 Neon (High reasoning)
52.3
%
1
Gemini
4 Carbon (High reasoning)
49.8
%
1
Claude
Sonnet 5.5 (Adaptive/Max)
47.7
%
1
Claude
Fable 5.1 (Adaptive/Max)
47.1
%
1
GPT
6.2 Astra (Max reasoning)
45.8
%
1
Gemini
4 Hydrogen (xHigh reasoning)
44.6
%
1
Grok
4.7 (xHigh reasoning)
44.0
%
1
Gemini
4 Argon (High reasoning)
36.6
%
1
GPT
6.2 Sol (Max reasoning)
34.5
%
1
Muse
Spark 1.3 (Max reasoning)
33.9
%
1
GPT
6 Astra (Max reasoning)
31.4
%
1
Qwen
3.8 Max (xHigh reasoning)
30.8
%
1
GLM
5.3 (Max reasoning)
29.9
%
1
Muse
Ember 1.0 (Max reasoning)
29.8
%
1
Muse
Wonder 1.0 (Max reasoning)
29.5
%
1
Kimi
K3 (Max reasoning)
28.6
%
1
GPT
6.1 Sol (Max reasoning)
28.6
%
1
Gemini
3.8 Flash (High reasoning)
26.5
%
1
Hy
Hy4 Preview (High reasoning)
24.3
%
1
DeepSeek
V4.1 Flash 1017 (Max reasoning)
24.0
%
1
Hy
Hy4 (High reasoning)
23.7
%
1
Inkling
2 Small (xHigh reasoning)
21.2
%
1
Mistral
Large 4 (High reasoning)
19.1
%
1
MiniMax
M3.1 Flash (Max reasoning)
13.8
%

Stay Posted