Benchmark
BankerToolBench
A banking-workflow benchmark used to evaluate agents and verifier behavior on artifact-heavy tasks.
Handshake · 100% completeConfirmed
Verifier·Public sample
Gandalf the Grader
An open-source reactive agent-as-judge that inspects files and environment state while grading agent work.
What it is
Gandalf runs inside a rollout environment and chooses what evidence to inspect using available tools. Handshake reports results on BankerVerifierBench and an internal productivity benchmark; those performance claims remain company-reported unless independently reproduced.
Capabilities
Only evidenced capabilities are shown as Yes. Missing evidence remains Unknown.
StatefulYes
Long horizonUnknown
Multi-turnUnknown
Multi-agentUnknown
AdversarialUnknown
MultimodalUnknown
Tool useYes
Browser useUnknown
Computer useUnknown
Code executionUnknown
File manipulationYes
API interactionUnknown
MCP supportUnknown
SandboxedUnknown
Deterministic resetUnknown
Stochastic behaviorUnknown
Human in the loopYes
Expert authoredUnknown
SyntheticUnknown
Production derivedUnknown
Real application replicaUnknown
Partial observabilityUnknown
Persistent stateUnknown
Cross-applicationUnknown
Sources & provenance
- Handshake AI Research · official · accessed 2026-08-16 · company-reported
- Handshake AI Research · repository · accessed 2026-08-16 · confirmed
Related artifacts
Public sample
RL Environment
RL Environments and Agents
Custom RL environments and verifier design for training and evaluating agentic models.
Surge · 100% completeCompany-reported
Dataset
Expert Professional Domains
Expert-authored data and judgment across professional, STEM, and humanities domains.
Surge · 100% completeCompany-reported