Gemini Agent vs Hermes Agent (NEW Google Interactions API)

Google promoted the Interactions API to its primary Gemini agent interface this week, signaling that the harness — not the model — is now where the value lives.

What Google shipped

The Interactions API is generally available as the new default endpoint for both Gemini models and agents, confirmed by Logan Kilpatrick. The feature set Phil Schmidt laid out: background execution, async execution, expanded tool support, multimodal generation, managed agents, and antfarm — an isolated remote Linux sandbox baked into the stack. Instead of an agent dumping files into your ~/Downloads, it executes in a clean workspace, which is why developers can finally let an LLM run code without bricking a local machine.

Google also shipped installable Gemini interaction skills — essentially a training manual that teaches coding agents the new SDK patterns and current model versions. The unspoken message: the hard part of new infrastructure is migration, not the spec.

What open source shipped

Hermes Agent's latest update adds iMessage access without a Mac, Raft integration as an external agent in shared workspaces, and GUI control for Windows and Linux desktop apps. The project just crossed 200K stars on GitHub — a developer signal that harness ergonomics are being voted on, not model benchmarks. Caveat: when running computer use, Hermes still guesses at UI element locations.

The unsolved problem

A recent survey of nine open-source agent communication protocols shows consensus on stateful memory and rich hybrid payloads, but no agreement on agent discovery across decentralized systems. Translation: cross-device task handoff (personal MacBook → work MacBook → Windows rig) is still the bottleneck for multi-agent workflows.

The framing

Google is selling the complete car — engine, chassis, garage. Hermes is shipping a hackable kit. Pick based on whether you want a polished vertical stack or cross-platform DIY control.