How Theo works

Figure 2. Theo’s research loop: ask, route, execute, preserve, review, and iterate.

The diagram represents a repeating control loop, not a rigid pipeline. You set the scientific intent and decide what happens next; Theo assembles the relevant research context, selects an execution path, and returns inspectable work for your review.

The five-stage loop

1. Intake — Establish the research focus, goals, preferences, context, and constraints for the session.

2. Research question or task — Ask in Central Chat, continue an existing chat, or manually choose a phase.

3. Orchestration — Theo considers the current chat, session goal and context, shared sources, and relevant artifacts, then chooses the best route.

4. Execution — The selected model or service performs the work, possibly using agents, subagents, scientific tools or a call to a phase.

5. Response and iteration — Review the answer and supporting work, continue in chat, refine or rerun a phase, branch to another approach, or revisit prior history.

Three possible routes

• Direct answer — For questions a model can answer well within the conversation.

• Scientific phase — For structured research tasks. Each phase is a FirstPrinciples service with its own workflow, agents, tools, and output format; results appear in the Phase UI.

• Council — For difficult questions that benefit from debate, critique, and synthesis across at least two models.

How state and provenance compound

The session keeps the research goal and shared context. Sources are attached to the session and can support multiple chats. Each chat preserves its own conversation and launches its own phase work. Artifacts remain associated with the chat and phase that produced them.

Phase outputs are saved as versioned records rather than silently overwritten. Refining the same phase creates a new artifact version; starting a different topic or research branch creates a separate phase with its own history.

Successful primary artifact versions are added to Lineage in build order. This lets you inspect provenance, see which inputs fed a result, return to earlier work, and follow branch outcomes through the chats and artifact versions that produced them.

Behind the interface, Theo records the requested work, specialized background services execute it, and completed results are written back to the same session. Theo surfaces the outcome in chat and returns control to you instead of automatically starting another phase unless you already requested the larger sequence.

What the researcher controls

You can choose the model or phase, provide or exclude sources, review assumptions, accept or reject results, refine a phase to create a new version, open a separate chat for another line of inquiry, use Council, and decide what should happen next.