Skip to main content

Modes

The mode buttons in the AI panel header decide how much latitude the assistant has. Switch between Plan and Investigate at any point in a conversation; Autoresearch must be chosen before the first message of a new chat.

Plan​

Plan mode is read-only. The assistant keeps every tool that looks at things - search_board, read_node, web_search, search_wikipedia, fetch_social_post, the court, SEC, sanctions and archive tools - and loses every tool that changes things. It can still ask you questions with ask_user.

Use Plan to:

  • Ask about the board. "Which actors have no connections?" "Summarize what we know about the Cyprus entities." "What's the weakest-sourced event here?"
  • Scope a job before spending tokens on it. "How would you approach mapping the ownership chain? Which sources would you use?" Agree the plan, then flip to Investigate and say "go".
  • Do research you want to read rather than file - a quick lookup you'll decide about later.

Because nothing can change, Plan is also the safe mode for exploring what a new provider or model does with your prompts.

Investigate​

The default. Everything Plan can do, plus creation, update and deletion of nodes, connections, entities and narratives. The assistant acts on your request directly: there is no per-action approval step. The safeguards are that it searches before creating, prefers updating an existing node to making a duplicate, is told to delete only for duplicates and corrections, and that everything it does is undoable with Ctrl/Cmd+Z and visible in the chat log.

If you want a review gate, use the board itself: ask it to put new material inside a Region called To review, or tag new nodes unverified, and clear them as you check them.

Autoresearch​

Autoresearch is an autonomous loop for open-ended investigation of a topic.

  1. Start a new chat, select Autoresearch, and describe the objective. Be as specific as you would be with a junior analyst: what you want to know, what counts as done, what to avoid.
  2. The assistant runs a normal Investigate turn.
  3. When it finishes, a supervisor pass - run on your summarizer model - reads the round and writes the next directive. The supervisor is deliberately adversarial: it looks for gaps and unverified claims, questions what has been taken at face value, points to loose threads, pushes for primary over secondary sources, and names specific searches to run next.
  4. The directive is fed back as the next user message and another round runs.
  5. This continues until the supervisor judges the investigation complete or 15 rounds have run, whichever is first. You can stop it at any time with the stop button.

What to expect: a well-scoped Autoresearch session produces a board with dozens of sourced nodes, connections and usually a couple of narratives, in the time it takes to make coffee. It also spends tokens accordingly - each round is a full turn - so watch the usage indicator if you are on the DeepState AI budget, and consider a cheaper model for very broad topics.

Tips:

  • Give it constraints. "Only use primary sources and filings", "stay within 2019-2023", "do not create predictions".
  • Give it a board with a few seed nodes rather than an empty one; it builds on what exists.
  • Review the supervisor directives afterwards. They are a good record of what the investigation considered and can be pasted into a Note as an audit trail.
  • Autoresearch is only available in fresh conversations because the supervisor needs to see the whole exchange. Once a chat has messages the button is disabled.

Reasoning and modes​

Reasoning depth is independent of mode. A useful pairing: Plan + high for strategy, **Investigate

  • low** for execution, Autoresearch + medium for the loop.