We want the Agents to be able to natively contribute to HM content.

User Stories

Translation

Summarization

Finding links

Critical Analysis

Idea generation

Importing knowledge

Community pays for tokens? Individuals pay for tokens?

Hide the use of my AI, keep context private

Publish the context and use of my AI, model metadata

Workflows

External Agent interface?

Claude Website

Skills with Claude Code

Internal Agent UX

Definitions

Agent

What is it for us?

System Prompt: Soul / Goal

Model Provider: Provider URL, API key, specify which model

Tools: what can the agent do

Which account, which types of mutations (read only? comment only? everything)

External tools like Web Search + Scrape

Access to other agents (compaction agent)

Schedule and Triggers: when does the model kick into action

Who owns an "agent"

Owned by a node? Because somebody needs to actually run the agent

Node has access to keys, private API keys

Needs some kind of administrator account or a node operator

Context

A Single chat history of an agent

Authorship

Sometimes the human wants to take responsibility for the agent's work. Sometimes the human wants to hide the use of AI and wants to take the credit from the model.

Other times, the human wants to provide full context about the creation of the content. Attribution to the agent, metadata that links to the context where the agent was prompted, and metadata about the model/time/cost/server.

@Eric Vicenti believes that both workflows are valid, but we should encourage giving the agent authorship, with the context and metadata shared transparently.

Option A: Per-Human-Account Agent Authors

asdf

Option B: Per-Agent Authors

A different author for each combination of

Option C: Human Authorship

Agents will use the human identities to sign.

Distributed Objects

Sync Agents and Context using our universal resources API, just like we want to do with drafts.

This allows collaboration between users in a community and syncing between device.

Principles

Flexibility

We want to provide the most powerful context management system as possible