Google models work for text completions but not for agent (tool-calling) benchmarks — the agent path raises
NotImplementedError (tjbench/models/anthropic_agent_client.py). Implement a ToolCallingClient for the
Google API (or route through the OpenAI-compatible layer if Google exposes a compatible surface) so Google
models can run the agent benchmarks with tool validation.
Acceptance criteria: a Google model completes sample-agent (or another agent benchmark) with the safety
gate / tool validation intact; the NotImplementedError is removed; tests cover the new client offline via
a mock.
Google models work for text completions but not for agent (tool-calling) benchmarks — the agent path raises
NotImplementedError(tjbench/models/anthropic_agent_client.py). Implement aToolCallingClientfor theGoogle API (or route through the OpenAI-compatible layer if Google exposes a compatible surface) so Google
models can run the agent benchmarks with tool validation.
Acceptance criteria: a Google model completes
sample-agent(or another agent benchmark) with the safetygate / tool validation intact; the
NotImplementedErroris removed; tests cover the new client offline viaa mock.