AI Agent

What it is for

Research workflows on FAVOR tend to look the same. You start with a question, pick a variant list or a gene, filter, rank, run a statistic, make a chart, and write it up. Most of those steps are mechanical once you know what you are trying to do. The agent exists to take a plain language question and walk that whole path for you, so you can read the result and decide what to do next.

The loop

Every turn follows the same shape. Think, do, check, repeat, or answer.

1. Read the workspace. What variant list is loaded, what fields are available, what did previous turns save. The agent starts every turn with a fresh snapshot of the session state, so it can make decisions based on the most up to date information.
2. Think. Based on the question and the workspace, the agent thinks of a few possible next steps it could take.
3. Pick a tool. Based on the question, the agent picks one action to take next.
4. Run it. The tool executes against real data. The result comes back as structured output.
5. Check. Did it work. Did it return nothing. Is something obviously wrong. If so, try a safer approach. If not, decide whether to stop or continue.
6. Synthesise. When enough steps have run, the agent writes a short answer grounded in the tool outputs and points at the artifacts (charts, tables, saved variant lists).

Tool surface

The agent can only do what its tools let it do. The tools are kept deliberately small, and each one has a single job.

Self correction

Most agent failures fall into a few categories. The system tries to handle each one before asking the user.

• Typos in field names. If the agent asks forcadd on a variant list that has cadd_phred, it gets automatically corrected silently.
• Empty results. When a query returns nothing, the pipeline probes the filters one at a time to figure out which one was the culprit, and suggests relaxing it.
• Degenerate results. If the agent runs something that returns a result but it is not useful (a chart with one bar, a gene list with one gene), it tries again with a more conservative approach.
• Method that cannot run. If the requested analysis cannot converge, the system falls back to a safer method and explains why.
• Repeated failures. If the same thing fails twice, the agent stops retrying and asks the user one clear question. It never loops forever.

When it asks

The agent asks the user in three situations.

1. Two good interpretations. You asked for something that could mean more than one thing, and both answers are valid.
2. The data cannot answer the question. The target has no variation, the variant list is too small, or the required columns are not present.
3. It is about to do something expensive. A query that would cost real resources is surfaced before it runs, so you can say yes or narrow it down.

Memory and sessions

The agent keeps three kinds of state.

• Session state. The current conversation, the active variant list, running jobs. Fresh at the start of every session.
• Pinned items. Entities you asked it to remember within a session. Used to avoid re resolving things you just talked about.
• Saved findings. Notes saved across sessions. The agent can look these up when a new question touches something you worked on before.

Charts

Charts are produced by deterministic mappers over tool outputs. The same output always makes the same chart.

Every chart is traceable to the tool call that produced it.