Plan mode and agent mode

The assistant runs in two modes, and they do not differ merely in convenience. In plan mode the model produces a workflow that a human reviews and the engine executes: the result is replayable. In agent mode the model calls the tasks directly: the exchange is more direct, and the resulting sequence is not reproducible.

Comparison

	Plan mode	Agent mode
Model output	A JSON object describing a workflow	Successive tool calls
Execution	By the engine, after human validation	Immediate, as the conversation goes
Replayable	Yes: the workflow is an artefact you can version, sign and hand over	No: the same prompt yields a different sequence
Traced	Yes	Yes
Writes	After the steps have been reviewed	Announced before execution, but decided by the model
Use	Regulated contexts, client deliverables, repeated procedures	Exploration, iteration, one-off questions

A language model is not deterministic: the same question can produce a different sequence of calls. The journal records what was done, but does not replay it identically. Plan mode covers the second need by producing a workflow, which can be replayed.

Plan mode

The model touches nothing. It receives the list of tasks and answers with a JSON object only, with no surrounding text.

{ "nom": "Synthèse Cr par lithologie", "etapes": [ { "id": "e1", "tache": "core.sql", "args": { "requete": "SELECT id, Cr, material_type FROM samples WHERE campaign='GM-2026'" } }, { "id": "e2", "tache": "core.resume", "args": { "element": "Cr", "par": "material_type" } }, { "id": "e3", "tache": "core.graphique", "args": { "type": "bar", "donnees": "${e2.resume}", "y": "moyenne" } } ] }

The plan is shown in the Assistant panel as readable steps. The user reviews it, corrects it if needed, then runs it. The validated workflow can be saved, exported and replayed from the Workflows panel.

Question › JSON plan › Human review › Executed by the engine › Workflow archived

Agent mode

The model calls tools as the conversation goes, up to a maximum of twelve turns. Every call is shown with its arguments, and every result is rendered on screen.

The system instruction requires the model to explain any action that writes before executing it. That is not a technical guarantee: the guarantee comes from the fact that the available tools delete nothing and never leave the machine.

Guard	Nature
The model explains before writing	An instruction. Not binding.
No task deletes anything	A property of the code. Binding.
No access to the DOM, the disk, the network	A property of the code. Binding.
Twelve tool turns maximum	A property of the code. Binding.
Every write recorded in the signed registry	A property of the code. Binding.

The distinction between the two columns matters for an audit. A prompt instruction describes expected behaviour and may not be followed; a property of the code applies on every call, whatever model is installed.

Streaming

The answer text arrives as a stream (NDJSON read line by line, with buffering across cut lines). Streaming is switched off when tools are offered to the model: a tool call arrives in fragments, and the model writes it as text. Reassembling it on the fly would mean guessing, at every fragment, whether you are in the middle of a JSON object or of a sentence, at the risk of displaying half a tool call.

In practice the tool-calling turn is short; it is the final answer turn that is long, and that one is streamed.

Choosing a mode

Situation	Mode
Exploring an unfamiliar dataset	Agent
Producing a figure for a manuscript	Plan, then archive the workflow
A procedure repeated on every campaign	Plan, workflow saved and versioned
A deliverable for an audited client	Plan. The workflow goes with the report.
A one-off question about a sample	Agent

A common method is to explore in agent mode, then ask for the matching plan once the approach has settled. The resulting workflow becomes the reproducible version of what was found by exploration.