The Biology Agent¶

The Biology Agent is an AI assistant that works on a model with you. It reads the model, searches the primary literature, and proposes concrete edits: add this state, add this process, this metabolite is an inhibitor of that enzyme, this rate law should be reversible. It can also set up the quantitative side, describe your experiments, and check a simulation against your data.

It is a collaborator, not an oracle. Every change it makes is one you approve, every change is visible in the diagram before you keep it, and nothing reaches the database until you commit.

Starting a session¶

Open a model and press AI Agent in the diagram toolbar. The Biology Agent window opens with the model already loaded. Describe what you want in your own words:

Check whether the regulation of PFK1 in this model matches the current literature, and add anything important that is missing.

This model has no feedback on the first committed step. Is there any, and if so add it.

Set up a Model Realization with modular rate laws and a MOE for my experiment 412, then simulate and tell me whether the shape is plausible.

Enter the time course I pasted below as a new experiment, attach it to MOE 4584, and tell me how far the current parameters are from it.

Press Send, or Enter (Shift+Enter starts a new line). The status line reports what it is doing — Planning…, Calling tool…, Waiting for your approval… — and the transcript fills in as it works. Stop ends the session at any point.

What it can do¶

The agent has a fixed set of tools and can do nothing outside them:

Read — the model's structure, molecules, places, processes, notes, parameter sets, and the results of the last simulation.

Search the web for primary sources, and cite them.

Read this documentation. When you ask how something in ProcessDB works, it looks the answer up here and cites the page. If the documentation does not cover it, it is instructed to say so rather than guess.

Change the model — add and remove states, processes, and modifier edges, and edit notes.

Build the quantitative side — create a Model Realization (seeded with defaults, as if you had done it), create a MOE, write rate laws, add MR and MOE functions, add events, and create parameter sets. It edits values only in sets it created itself: set 0 and the sets you made are never touched.

Work with experiments — read the experiments in the database, their protocols, measurements and data; create an experiment and attach it to a MOE; add measurements, data sets and data points; record protocols of every kind (bolus, infusion, events, graph, formula, forcing function); and say which model variable a measurement corresponds to. It is instructed never to invent data: what it enters is what you gave it, or what it read from a source it cites.

Compare simulation to data — for each data set, the residuals and the sum-of-squares and root-mean-square error, so it can judge its own proposal and tune parameters by hand. It has no optimizer; the optimizer stays yours to run.

Run things — regenerate a MOE's equations, run a simulation, and read the results back.

Ask you — when it needs a decision only you can make, it puts a question in the transcript with a few options to click and a box for anything else.

It works the way the modelling does: structure first, then quantification, then data.

Approval¶

Every tool call that changes something stops and asks. The Confirm Action dialog names the tool, summarises the action in a sentence, and shows the arguments. You can:

• Allow — run this one;
• Deny — skip it. The agent is told not to retry the same action, and adapts;
• Allow all this session — stop asking for the rest of this session.

Read-only calls — reading the model, searching the literature, reading this documentation — do not prompt. Running a simulation and regenerating equations do, since they change what is on screen.

When the agent finds an existing process that matches one it is about to create, it asks in the transcript whether to Reuse process or Create new process, with an always version of each for the rest of the session.

Reviewing what it did¶

When the session ends, a banner appears in the agent window and above the diagram: Agent changes: N added · M removed. Review the diagram, then accept or discard. Added items carry a green halo and removed items are drawn as red ghosts in place. Accept keeps the changes and clears the highlighting; Discard undoes every change the agent made, leaving your own edits from the same session alone.

So the real review point is not the approval dialog but the diagram: approve a plausible-sounding edit, look at what it actually did to the network, and discard it if you do not like it.

Discard is unavailable once something has been saved during the session — changes that have reached the database cannot be rolled back by the banner. Accept them and edit by hand instead.

And nothing is permanent regardless: agent changes are ordinary local edits. They reach the database only when you commit, and Undo works on them like anything else.

Past conversations¶

The ≡ Conversations button shows and hides the list of previous sessions on this model. Select one to read what was proposed and why — including the citations — which makes the agent's transcripts a usable record of where a model's claims came from. Typing a message while a past conversation is showing resumes it; + New Conversation starts afresh.

The toolbar also has Auto Scroll, a font-size control for the transcript, and a running token count. Each message has a Copy button that copies it with formatting intact, for pasting into a document.

TheBrain¶

If you keep notes in TheBrain, the agent can read them as background — your own accumulated knowledge alongside the published literature. The first time it tries, a TheBrain access dialog asks your permission separately from the agent's own approval gate, and ProcessDB remembers your answer. Connecting TheBrain in the first place is offered when ProcessDB starts, and walks you through enabling TheBrain's Local API.

Getting good results¶

• It works best on one question at a time. "Review the regulation of this enzyme" gets a sharper answer than "improve this model".
• It follows your framing. Saying which parts of the model you are already confident in keeps it working where you want it.
• Its sources are in the transcript. The agent is instructed to cite reliably and not to invent references, and every citation it gives is there to be traced.
• The quantitative side suits it too. Setting up a realization, seeding rate laws, entering an experiment, running a simulation and reporting the residuals is the kind of mechanical work it does well.
• It is told to fit the data, not defeat it. Lowering an error by down-weighting inconvenient points or rebinding a measurement to a better-behaved variable is not a result, and the agent is instructed to say which residual drove each change it makes.
• The biology stays yours to judge. The agent is good at finding what the literature says and turning it into model structure. Whether that belongs in your hypothesis is a different question, and the approval gate and the diff in the diagram exist so ProcessDB never answers it for you.

Exporting a model to another AI tool¶

For reasoning about a model in a chat tool of your own, File → Export JSON (for LLMs) writes the current diagram as a file you can attach to a conversation. Nothing is sent anywhere by that command; it just produces a file.