Research plans

What a research plan is

A research plan is a structured document the agent builds to organize a multi-step investigation. It contains:

• Goal: what you're trying to find out
• Hypotheses: ranked guesses about what will work
• Experiments: concrete runs to test each hypothesis
• Success criteria: how you'll know if a hypothesis is confirmed
• Dependencies: what needs to happen before what

Think of it as a decision tree for your research. Instead of running one experiment at a time and deciding what's next, the agent maps out the whole plan upfront.

How the agent builds one

Tell the agent something like:

I want to find the best learning rate schedule for fine-tuning Llama 3 8B on my dataset. Build a research plan.

The agent will:

1. Review your dataset and model configuration
2. Read relevant papers and best practices
3. Generate a plan with 3 to 8 experiments
4. Present it for your review

The plan appears in the center panel as an interactive tree.

Approving and modifying

You can:

• Approve the whole plan. The agent starts executing experiments in order
• Modify individual steps. Change hyperparameters, add experiments, remove ones you don't care about
• Reorder priorities. Drag experiments to change execution order
• Add a branch. If a hypothesis suggests a follow-up, add it to the tree

Nothing runs until you approve. You stay in control.

The autonomous loop

Once approved, the agent executes the plan autonomously:

1. Runs the next experiment
2. Collects results
3. Updates the hypothesis status (confirmed / rejected / inconclusive)
4. Picks the next experiment based on results
5. Repeats until the plan is complete or you intervene

You can pause, modify, or cancel the plan at any time. The agent sends you a notification when the plan completes or when it needs your input on an ambiguous result.