Hermes Agent 0.22: the first minutes decide adoption

On October 13, Nous Research released Hermes Agent v0.22.0 with the subtitle The Threshold Release.

After a fresh install, the user enters a setup chat where the agent guides them step by step.

The chat uses a free trial model, so newcomers talk to the agent before creating an API key.

For a company adopting AI, this suggests a practical rule: the first minutes determine an agent’s fate, and they deserve most of the engineering attention.

Nous Research’s release notes quantify the work between v0.21.0 and v0.22.0: about 24,200 commits and 6,080 merged pull requests; 15,600 changed files; roughly 2.05 million added and 1.14 million deleted lines; 7,180 closed issues; and 2,520 contributors.

The ratio is particularly interesting: 56 lines were deleted for every 100 added.

We interpret these figures as substantial rewriting alongside new features. The release notes do not explicitly say so; this is our interpretation.

The authors highlight two user-facing changes: guided setup in a chat using a trial model, and connectors letting the agent operate inside familiar applications.

At KT.Team, we evaluate tools using TTU, time to use: time from installation to the first useful result.

The typical open-source agent setup has five steps: install it, register with a model provider, add a payment card, create a key and enter it in configuration.

Only then does the user receive a first answer.

Some people abandon the attempt at every step.

We have not seen Nous’s funnel data, so specific drop-off rates remain our hypothesis.

We have seen a different pattern in clients’ internal products: employees put off tools requiring several actions before a first result and never return. Nous removed the most expensive first step: obtaining a key and paying.

It looks like a convenience improvement, but the team has changed the order of the transaction.

Users first experience the agent’s value and only then pay.

Hermes connectors do the same job as corporate integrations: they let the agent read and change data where people work. Corporate systems use the MCP standard, Model Context Protocol, for this.

The agent uses it to call external tools: find an order in 1C, update an attribute in Pimcore or Akeneo, or check an electronic document’s status. In AI-native integration projects, we assemble four components: - MCP servers in front of accounting systems. Every operation has an explicit contract, permissions and an audit log.

The agent sees only permitted operations, without gaining access to the entire database. - A message bus between the agent and accounting.

Requests pass through Apache Kafka or Datareon.

The bus absorbs spikes in agent requests, keeping 1C available during working hours. - LLM & Security Gateway.

The gateway routes requests to Claude, GigaChat, YandexGPT or Qwen, masks personal data before external API calls and records every call for audit. - RAG over internal documents.

Policies and contracts enter the model context with a link to